remote/datasette
1
0
Fork 0
mirror of https://github.com/simonw/datasette.git synced 2026-09-27 04:14:25 +02:00
Code Issues Projects Releases Packages Wiki Activity

Simplify code comments added since 1.0a40 (#2957)

Browse source 
Refs #2867
This commit is contained in:
Simon Willison 2026-09-24 13:52:29 -07:00 • committed by GitHub
commit 8e17729ff3
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
14 changed files with 412 additions and 1451 deletions
Download patch file Download diff file Expand all files Collapse all files

9
tests/conftest.py
View file

@ -58,8 +58,6 @@ def find_free_port():
return sock.getsockname()[1]
# The otel fixtures moved to datasette.telemetry_testing, which is public
# plugin API - core's suite consumes it exactly the way a plugin's would.
from datasette.telemetry_testing import ( # noqa: F401
MetricsCollector,
otel_meter_provider,
@ -183,11 +181,8 @@ def pytest_collection_modifyitems(config, items):
move_to_front(items, "test_spatialite_error_if_attempt_to_open_spatialite")
move_to_front(items, "test_package")
move_to_front(items, "test_package_with_port")
# Same reason: this one shells out to a fresh interpreter. Late in a serial
# run the pytest process holds enough threads that the fork half of
# subprocess' fork+exec crashes the interpreter on macOS/CPython 3.13
# (SIGSEGV/SIGBUS inside _execute_child). Reproduces with any subprocess
# call placed there, on an unmodified tree - running it first avoids it.
# These start subprocesses, which can crash on macOS/CPython 3.13 late in
# a test run once the pytest process has started many threads
move_to_front(items, "test_datasette_package_never_imports_the_sdk")
move_to_front(items, "test_kit_module_itself_never_imports_the_sdk")
move_to_front(items, "test_no_provider_takes_the_fast_path")

286
tests/test_http_span.py
View file

@ -1,23 +1,6 @@
"""
The HTTP request span.
`tests/test_telemetry_registry.py` already pins the span's name shape, kind
and attribute keys against literals, so this file deliberately does not
repeat that. What it covers is the properties of the middleware and of the
router's `http.route` enrichment that the registry conformance test
structurally cannot see:
- **where the middleware sits.** Outermost is the entire point - moving it
inside the plugin `asgi_wrapper()` loop leaves plugin middleware creating
orphan root traces, which is the problem this span exists to fix, and every
attribute assertion still passes.
- **which span the route lands on**, which only diverges once something else
has made a span current.
- **method clamping**, which a workload of ordinary GETs can never exercise.
- **the query string never being recorded**, which only fails if a request
actually carries one.
- **the span outliving a streamed response body**, which only a paging export
can distinguish from ending far too early.
Tests for the HTTP request span created by TelemetryMiddleware and the
`http.route` enrichment added by the router.
"""
import asyncio
@ -51,9 +34,8 @@ from datasette.telemetry import (
)
from datasette.utils import resolve_routes
# Named in-memory databases are shared-cache: two Datasette instances given
# the same name share one SQLite database and the second `create table`
# fails.
# Named in-memory databases are shared between instances, so each fixture
# needs a unique name.
_names = itertools.count()
@ -61,7 +43,7 @@ PLUGIN_MIDDLEWARE_SPAN = "test.plugin.middleware"
class _MiddlewarePlugin:
"A plugin asgi_wrapper() that creates a span, standing in for a real one."
"A plugin asgi_wrapper() that creates a span."
__name__ = "HttpSpanMiddlewarePlugin"
@ -79,11 +61,8 @@ class _MiddlewarePlugin:
class _RaisingMiddlewarePlugin:
"""
A plugin asgi_wrapper() that raises.
`route_path` converts almost every exception into a 500 itself, so an
exception escaping into the request span is only reachable from *outside*
the router - a plugin wrapper, or a failure inside the 500 handler.
A plugin asgi_wrapper() that raises. `route_path` turns most exceptions
into a 500, so this is how an exception reaches the request span.
"""
__name__ = "HttpSpanRaisingMiddlewarePlugin"
@ -135,18 +114,12 @@ async def ds():
@pytest_asyncio.fixture
async def ds_paging():
"""
An instance whose table is bigger than `max_returned_rows`.
That is what makes `?_stream=1` genuinely page: `stream_csv` loops calling
`fetch_data` for each page *inside* the response body send, so the trace
contains `db.query` spans that start after the response has begun. On a
table that fits in one page every query finishes before the body starts
and the span-covers-the-body assertion cannot fail.
An instance whose table is bigger than `max_returned_rows`, so a
`?_stream=1` export runs queries for later pages during the body send.
"""
name = f"httpspanpaging{next(_names)}"
# Both settings matter. `?_stream=1` forces `_size=max`, which is
# `max_returned_rows` - so lowering only that gives one page of five rows
# and no `next` token, and the export never loops.
# Both settings are needed: lowering only max_returned_rows gives a
# single page with no `next` token.
instance = Datasette(
memory=True, settings={"max_returned_rows": 5, "default_page_size": 3}
)
@ -182,13 +155,8 @@ async def test_plugin_asgi_wrapper_middleware_runs_inside_the_request_span(
ds, otel_spans
):
"""
The placement check.
A span created by a plugin `asgi_wrapper()` must be a *child* of the
request span. If the middleware is mounted anywhere inside the plugin
loop the two swap places - the plugin's span becomes the root and the
request span its child - which is exactly the orphaning this is meant to
prevent, and which no attribute assertion notices.
Spans created by plugin asgi_wrapper() middleware are children of the
request span.
"""
ds.pm.register(_MiddlewarePlugin(), name="httpspan-middleware")
try:
@ -210,7 +178,7 @@ async def test_plugin_asgi_wrapper_middleware_runs_inside_the_request_span(
assert plugin_spans[0].parent.span_id == server_span.context.span_id
assert plugin_spans[0].context.trace_id == server_span.context.trace_id
# And the database work is in the same trace, not off on its own.
# Database spans are in the same trace.
queries = [span for span in spans if span.name == "db.query"]
assert queries, "a table page should have issued at least one query"
for query in queries:
@ -220,15 +188,8 @@ async def test_plugin_asgi_wrapper_middleware_runs_inside_the_request_span(
@pytest.mark.asyncio
async def test_unrecognised_method_is_clamped(ds, otel_spans):
"""
Anyone can send `FROB / HTTP/1.1`. An unclamped method is an unbounded
dimension a client controls, so semantic conventions map anything off the
known list to `_OTHER`.
The span name is checked too, and it is the reason the router clamps the
method a second time when it renames the span: the middleware's clamping
protects the attribute, but the name is rebuilt from `request.method` in
`route_path`, which is the raw client string. An unclamped rename would
put attacker-supplied text straight back into the span name.
Unknown methods are recorded as `_OTHER` in both the attribute and the
span name, which the router rebuilds from the raw `request.method`.
"""
otel_spans.clear()
await ds.client.request("FROB", f"/{ds.db_name}/t")
@ -240,7 +201,7 @@ async def test_unrecognised_method_is_clamped(ds, otel_spans):
@pytest.mark.asyncio
async def test_known_method_is_not_clamped(ds, otel_spans):
"The other half of clamping: a real method must survive it verbatim."
"Known methods are recorded unchanged."
otel_spans.clear()
await ds.client.get(f"/{ds.db_name}/t")
server = _server_spans(otel_spans)
@ -251,12 +212,7 @@ async def test_known_method_is_not_clamped(ds, otel_spans):
@pytest.mark.asyncio
async def test_the_query_string_is_never_recorded(ds, otel_spans):
"""
Datasette puts user-supplied SQL in `?sql=` and canned query parameters in
the query string, so no span may carry it. Asserting on the absence of a
`url.query` key alone would not catch it arriving under some other name,
so this searches every attribute value of every span for the marker.
"""
"No attribute on any span contains the query string."
marker = "canary-9f2b1c"
otel_spans.clear()
await ds.client.get(f"/{ds.db_name}/t?_facet=v&_nosuch={marker}")
@ -283,9 +239,8 @@ async def test_url_path_is_recorded_without_the_query_string(ds, otel_spans):
@pytest.mark.asyncio
async def test_escaping_exception_sets_error_type_and_reraises(ds, otel_spans):
"""
An exception that gets past `route_path` must be recorded, not swallowed.
No response ever started, so there is no status code to record either.
An exception that escapes `route_path` is recorded and re-raised. No
response started, so no status code is recorded.
"""
ds.pm.register(
_RaisingMiddlewarePlugin(call_app_first=False), name="httpspan-raiser"
@ -308,11 +263,8 @@ async def test_an_escaping_exception_beats_the_status_code_for_error_type(
ds, otel_spans
):
"""
Both paths can fire on one request: a 500 response is sent and *then*
something raises on the way out. The `finally` block runs while the
exception is propagating, so without the guard it would overwrite the
exception's class name with the string "500" - strictly less information
about what actually went wrong.
A 500 response followed by an exception records the exception class as
`error.type`, not "500".
"""
ds.pm.register(_BoomPlugin(), name="httpspan-boom")
ds.pm.register(
@ -327,24 +279,16 @@ async def test_an_escaping_exception_beats_the_status_code_for_error_type(
ds.pm.unregister(name="httpspan-boom")
server = _server_spans(otel_spans)
assert len(server) == 1
# The 500 really was sent, so the status is still recorded ...
assert server[0].attributes["http.response.status_code"] == 500
# ... but error.type names the exception, not the status.
assert server[0].attributes["error.type"] == "RuntimeError"
@pytest.mark.asyncio
async def test_a_404_is_not_an_error(ds, otel_spans):
"""
Per semantic conventions a 4xx is the client's mistake, not the server's,
so a SERVER span must record the status and leave both its own status and
`error.type` alone. Datasette 404s are routine - every missing table, and
every bot probing for /wp-login.php - so treating them as errors would
drown a real 500 in noise.
Note this 404 *does* match a route: `/no-such-database-at-all` matches the
database pattern and the view then raises `NotFound`. Most Datasette 404s
are that shape rather than the unrouted one below.
A 4xx records the status code but no `error.type` or error status.
`/no-such-database-at-all` matches the database route, so `http.route`
is still set.
"""
otel_spans.clear()
response = await ds.client.get("/no-such-database-at-all")
@ -354,7 +298,6 @@ async def test_a_404_is_not_an_error(ds, otel_spans):
assert server[0].attributes["http.response.status_code"] == 404
assert "error.type" not in server[0].attributes
assert server[0].status.status_code is StatusCode.UNSET
# Route enrichment must not be gated on a successful response.
assert "http.route" in server[0].attributes
assert server[0].name != "GET"
@ -362,14 +305,8 @@ async def test_a_404_is_not_an_error(ds, otel_spans):
@pytest.mark.asyncio
async def test_an_unrouted_404_has_no_route_and_a_bare_method_name(ds, otel_spans):
"""
When no route matches there is nothing to set `http.route` to, so the span
keeps the bare method name it was given at the edge - which is exactly the
fallback semantic conventions specify for an unknown route.
`/a/b/c/d/e` is used rather than a plausible-looking missing name because
Datasette's route table is greedy: `/no-such-database-at-all` matches the
database pattern, and `/-/nope/deeper` matches the row pattern. Only a
path deeper than any route matches nothing at all.
With no matching route the span keeps the bare method name. Most missing
paths still match a route, so this uses a path deeper than any route.
"""
otel_spans.clear()
response = await ds.client.get("/a/b/c/d/e")
@ -384,14 +321,7 @@ async def test_an_unrouted_404_has_no_route_and_a_bare_method_name(ds, otel_span
@pytest.mark.asyncio
async def test_only_the_first_http_response_start_is_recorded(otel_spans):
"""
The `send` wrapper keeps the first status it sees.
Nothing in Datasette sends two `http.response.start` messages, so this
drives the middleware directly rather than pretending a request could
reach it. Without the guard a misbehaving plugin's second start message
would silently replace the status the client actually received.
"""
"The `send` wrapper records the status from the first `http.response.start`."
async def two_starts(scope, receive, send):
await send({"type": "http.response.start", "status": 200, "headers": []})
@ -418,10 +348,8 @@ async def test_only_the_first_http_response_start_is_recorded(otel_spans):
@pytest.mark.asyncio
async def test_lifespan_scope_passes_through_unspanned(otel_spans):
"""
`AsgiLifespan` sits *inside* this middleware, so the scope-type check has
to come first or startup and shutdown events never reach it. A SERVER
span for a lifespan scope is the symptom of that check being missing or
late.
Lifespan scopes reach `AsgiLifespan`, which sits inside this middleware,
without creating a SERVER span.
"""
instance = Datasette(memory=True)
app = instance.app()
@ -442,13 +370,7 @@ async def test_lifespan_scope_passes_through_unspanned(otel_spans):
@pytest.mark.asyncio
async def test_http_route_is_the_compiled_pattern(ds, otel_spans):
"""
`http.route` is the route's compiled regex, not a prettified template.
Asserted against what Datasette's own router resolves rather than against
a copied literal, so this pins the *relationship* - the attribute is the
matched route - and does not break when a core pattern is edited.
"""
"`http.route` is the compiled regex of the route Datasette's router resolves."
path = f"/{ds.db_name}/t"
expected = _route_for(ds, path)
otel_spans.clear()
@ -457,9 +379,7 @@ async def test_http_route_is_the_compiled_pattern(ds, otel_spans):
assert len(server) == 1
assert server[0].attributes["http.route"] == expected
assert server[0].name == f"GET {expected}"
# The pattern really is the ugly one, and that is deliberate - if someone
# adds a prettifier this is the assertion that should make them argue for
# it rather than slip it in.
# The raw pattern, not a prettified template:
assert "(?P<database>" in expected
@ -469,16 +389,8 @@ async def test_the_route_lands_on_the_request_span_not_a_plugins_current_span(
):
"""
The route is set on the span the middleware started, found through the
ASGI scope - not on whatever span happens to be current when routing
resolves.
Those are the same span only until a plugin `asgi_wrapper()` starts one of
its own. A plugin wrapper runs *inside* this middleware, so an instrumented
plugin makes its span current for the whole request: reading the current
span in `route_path` renames that plugin's INTERNAL span to
`GET <route>` and hangs `http.route` off it, while the actual request span
keeps a bare method name and never gets the one attribute a trace UI
groups requests by. Verified by reproducing it, not by reasoning about it.
ASGI scope, not on a plugin `asgi_wrapper()` span that is current during
routing.
"""
ds.pm.register(_MiddlewarePlugin(), name="httpspan-middleware")
try:
@ -494,7 +406,7 @@ async def test_the_route_lands_on_the_request_span_not_a_plugins_current_span(
assert len(server) == 1
assert server[0].attributes["http.route"] == expected
assert server[0].name == f"GET {expected}"
# And the plugin's span is untouched: same name, no route attribute.
# The plugin's span keeps its name and has no route attribute.
plugin_spans = [span for span in spans if span.name == PLUGIN_MIDDLEWARE_SPAN]
assert len(plugin_spans) == 1
assert "http.route" not in (plugin_spans[0].attributes or {})
@ -502,7 +414,7 @@ async def test_the_route_lands_on_the_request_span_not_a_plugins_current_span(
@pytest.mark.asyncio
async def test_request_span_attributes(ds, otel_spans):
"The whole attribute set on one ordinary request."
"The attributes recorded for an ordinary request."
path = f"/{ds.db_name}/t"
otel_spans.clear()
assert (await ds.client.get(path)).status_code == 200
@ -515,8 +427,7 @@ async def test_request_span_attributes(ds, otel_spans):
assert attributes["http.response.status_code"] == 200
assert attributes["http.route"] == _route_for(ds, path)
assert server[0].status.status_code is StatusCode.UNSET
# Never, on any span: an IP is borderline PII and the query string carries
# user-supplied SQL.
# The client IP address and query string are not recorded.
assert "client.address" not in attributes
assert "url.query" not in attributes
@ -524,12 +435,8 @@ async def test_request_span_attributes(ds, otel_spans):
@pytest.mark.asyncio
async def test_db_query_spans_are_children_of_the_request_span(ds, otel_spans):
"""
The point of the whole PR.
Not just "same trace ID" - every `db.query` span must reach the request
span by walking parents, and the request span must be the only root. A
stray root would show up in a trace UI as its own single-span trace, which
is the state this replaces.
Every `db.query` span descends from the request span, which is the only
root span.
"""
otel_spans.clear()
assert (await ds.client.get(f"/{ds.db_name}/t?_facet=v")).status_code == 200
@ -550,7 +457,7 @@ async def test_db_query_spans_are_children_of_the_request_span(ds, otel_spans):
assert queries, "a faceted table page should have issued queries"
for query in queries:
assert query.context.trace_id == server_span.context.trace_id
# Walk up to the root, which must be the request span.
# Walk up to the root, which should be the request span.
current = query
seen = 0
while current.parent is not None:
@ -563,9 +470,8 @@ async def test_db_query_spans_are_children_of_the_request_span(ds, otel_spans):
@pytest.mark.asyncio
async def test_500_sets_error_status_and_error_type(ds, otel_spans):
"""
A plain 500 - no exception escaping the app, because `route_path` converts
it into a response itself. The status is the only signal the middleware
gets, so `error.type` is the status as a string.
`route_path` turns the exception into a 500 response, so `error.type` is
the status code as a string.
"""
ds.pm.register(_BoomPlugin(), name="httpspan-boom")
try:
@ -584,25 +490,11 @@ async def test_500_sets_error_status_and_error_type(ds, otel_spans):
@pytest.mark.asyncio
async def test_csv_stream_span_covers_the_body_send(ds_paging, otel_spans):
"""
The span must not end when the handler returns - it has to cover the
response body.
The request span covers a streamed CSV body, including queries for later
pages that run after the response has started.
`stream_csv` runs its generator inline inside `AsgiStream.asgi_send`, and
that call happens inside the single `await self.app(...)` the middleware
makes, so a plain `finally` is enough and no deferred-end machinery is
needed. This is the assertion that holds that claim up: a `db.query` that
starts during the body send must still finish before the request span
does.
Only meaningful on an export that actually pages, hence `ds_paging` - on a
single-page table every query is over before the body begins and this
passes however early the span ends. The middle assertion below, that some
query *started* after `http.response.start` went out, is what keeps the
test honest about that; it is why the app is driven as raw ASGI rather
than through `ds.client`, which cannot timestamp the response start.
`time.time_ns()` is the same clock the SDK stamps spans with, so the two
are directly comparable.
Driven as raw ASGI to timestamp `http.response.start` with `time.time_ns()`,
the clock the SDK uses for spans.
"""
app = ds_paging.app()
body = []
@ -634,7 +526,7 @@ async def test_csv_stream_span_covers_the_body_send(ds_paging, otel_spans):
receive,
send,
)
# 40 rows plus a header - the export really did read past one page
# 40 rows plus a header, so the export read past the first page
assert len(b"".join(body).decode("utf-8").strip().splitlines()) == 41
assert response_started_at is not None
@ -662,12 +554,8 @@ async def test_csv_stream_span_covers_the_body_send(ds_paging, otel_spans):
@pytest.mark.asyncio
async def test_inbound_traceparent_becomes_the_parent(ds, otel_spans):
"""
W3C trace context is extracted with the global propagator, so a request
from an already-traced caller continues that trace.
The sampled flag has to be set: the SDK's default sampler is
parentbased_always_on, so a `-00` flag would drop the span and the test
would fail for a reason that has nothing to do with propagation.
An inbound `traceparent` header continues the caller's trace. It uses the
sampled flag (`-01`) because the SDK's default sampler is parent-based.
"""
trace_id = "4bf92f3577b34da6a3ce929d0e0e4736"
parent_span_id = "00f067aa0ba902b7"
@ -684,7 +572,7 @@ async def test_inbound_traceparent_becomes_the_parent(ds, otel_spans):
assert server_span.parent is not None
assert f"{server_span.parent.span_id:016x}" == parent_span_id
assert server_span.parent.is_remote
# And the database spans joined the caller's trace too, not a new one.
# Database spans are in the caller's trace too.
queries = [
span for span in otel_spans.get_finished_spans() if span.name == "db.query"
]
@ -696,20 +584,12 @@ async def test_inbound_traceparent_becomes_the_parent(ds, otel_spans):
@pytest.mark.asyncio
async def test_user_supplied_sql_in_the_query_string_is_never_recorded(ds, otel_spans):
"""
The `?sql=` case specifically, which is the one that matters: this is the
request where the query string *is* user-supplied SQL, and it reaches a
view that runs it. The marker is searched for across every attribute of
every span in the trace, not just for a `url.query` key, so recording it
under some other name fails too.
`db.query.text` legitimately contains the SQL - that is documented and
deliberate - so the marker is checked against the request span's own
attributes, and against `url.*` and `http.*` keys everywhere.
SQL from `?sql=` is not recorded on the request span or in any `url.*`
or `http.*` attribute. `db.query.text` is expected to contain it.
"""
marker = "secret_marker_5b1f"
otel_spans.clear()
# `/{db}?sql=` 302s to the query view, so go straight there - a redirect
# would leave the SQL only on a span for a request that never ran it.
# `/{db}?sql=` redirects to the query view, so request that directly.
response = await ds.client.get(f"/{ds.db_name}/-/query?sql=select+'{marker}'")
assert response.status_code == 200
spans = otel_spans.get_finished_spans()
@ -723,26 +603,15 @@ async def test_user_supplied_sql_in_the_query_string_is_never_recorded(ds, otel_
and (marker in str(value) or str(key) == "url.query")
]
assert not leaked, "the query string reached a span attribute: " + ", ".join(leaked)
# The request really did carry the marker, so the search above had
# something to find.
# Confirm the query ran with the marker.
assert marker in response.text
def test_request_span_skips_a_valid_but_non_recording_span():
"""
`request_span()` is guarded on `is_recording()`, not on
`get_span_context().is_valid`, and this is the case that separates them.
With no provider installed but an inbound `traceparent`, the API's
NoOpTracer hands back a `NonRecordingSpan` carrying the *remote* span
context - valid, sampled, and recording nothing. An `is_valid` guard would
wave that through and the router would build the name string and call
`set_attribute`/`update_name` on a span that discards both.
Tested at this level deliberately: through a real request the two guards
are indistinguishable, because every call the router makes on a
NonRecordingSpan is already a no-op. The only difference is the work done
to get there, so the guard itself is what has to be asserted on.
`request_span()` returns None for a `NonRecordingSpan` with a valid remote
span context, which is what an inbound `traceparent` produces with no
provider installed.
"""
remote = SpanContext(
trace_id=0x4BF92F3577B34DA6A3CE929D0E0E4736,
@ -754,13 +623,13 @@ def test_request_span_skips_a_valid_but_non_recording_span():
non_recording = NonRecordingSpan(remote)
assert non_recording.is_recording() is False
assert request_span({REQUEST_SPAN_SCOPE_KEY: non_recording}) is None
# Nothing current, nothing in the scope: the INVALID_SPAN fallback.
# No span in the scope and no current span:
assert request_span({}) is None
# And the case it must not skip.
# A recording span is returned:
with tracer.start_as_current_span("test.request_span.recording") as span:
assert request_span({REQUEST_SPAN_SCOPE_KEY: span}) is span
# Falling back to the current span is how an externally installed
# SERVER span still gets enriched.
# Falls back to the current span, such as one created by another
# SERVER instrumentation:
assert request_span({}) is span
@ -820,26 +689,11 @@ NO_PROVIDER_PROGRAM = textwrap.dedent("""
def test_no_provider_takes_the_fast_path():
"""
With no `TracerProvider` installed the middleware must hand the
application the *original* `send`, not a wrapper - a default Datasette
install should pay essentially nothing for instrumentation it is not
using.
With no `TracerProvider` installed the middleware passes the original
`send` to the application, including for requests with a `traceparent`.
This has to run in a subprocess. The suite's `otel_provider` fixture is
session-scoped and autouse, and `set_tracer_provider()` is effectively
once-per-process, so in-process every span is recording and the fast path
is unreachable.
The second case, with an inbound `traceparent`, is the one that pins the
check itself. With no provider the API's NoOpTracer returns a
NonRecordingSpan carrying the *remote* span context: its
`get_span_context().is_valid` is True while `is_recording()` is False. A
fast path guarded on `is_valid` would therefore silently stop working for
exactly the requests that arrive from an already-traced caller - which on
a real deployment behind an instrumented proxy is all of them.
conftest.py's pytest_collection_modifyitems() moves this test to the front
of the run by name - if you rename it, rename it there too.
Runs in a subprocess because the suite installs a provider for the whole
process. conftest.py moves this test to the front of the run by name.
"""
result = subprocess.run(
[sys.executable, "-c", NO_PROVIDER_PROGRAM],
@ -854,19 +708,15 @@ def test_no_provider_takes_the_fast_path():
"entry is the inbound-traceparent case, which fails if the fast path "
"is guarded on is_valid instead of is_recording()"
)
# Same fast path, other observable: nothing is stashed in the scope either.
# Nothing is stored in the scope either.
assert report["scope_keys"] == [False, False]
@pytest.mark.asyncio
async def test_internal_client_requests_are_marked(ds, otel_spans):
"""
An in-process `datasette.client` request runs the full ASGI stack, so it
emits its own SERVER span - `datasette.internal_client` marks those so
kind-based dashboards can filter the double-count out. A request that
arrives through the raw ASGI app (the shape of a real inbound request,
without the DatasetteClient wrapper setting the ContextVar) must not
carry the attribute.
`datasette.internal_client` is set on SERVER spans for `datasette.client`
requests, but not for requests made directly to the ASGI app.
"""
otel_spans.clear()
assert (await ds.client.get("/")).status_code == 200

33
tests/test_internals_database.py
View file

@ -1329,27 +1329,11 @@ async def test_write_thread_context_is_detached_between_tasks(
tmp_path, monkeypatch, num_sql_threads
):
"""
The write thread attaches each task's otel Context and must detach it
again before picking up the next task. The thread is persistent and
shared, so a leaked token would grow that thread's context stack for the
rest of the process - and a *wrong*-token detach only logs a warning
rather than raising, so "does it throw" cannot catch either mistake.
The write thread attaches each task's OpenTelemetry context and detaches
it before the next task, including when the task raises an exception.
Two things are asserted, because neither alone is sufficient:
1. Each task observes the context value that was current on the event
loop when it was queued. This is what fails if the Context is not
carried on WriteTask, or is never attached. It does *not* catch a
missing detach: attach() replaces the current Context wholesale, so a
leftover one from a previous task is simply overwritten.
2. The write thread's attach depth is identical at the same point in
every task. This is what fails if detach is missing - the stack grows
by one per task - and it holds across a task that raises, because the
detach lives in a `finally`.
An otel context value is used rather than a plain contextvars.ContextVar:
a plain var set on the event loop never crosses into the write thread, so
the probe would read None every time and the test could not fail.
Checks that each task sees the context from when it was queued, and that
the write thread's attach depth does not grow between tasks.
"""
name = f"context_leak_test_{num_sql_threads}"
db_path = tmp_path / f"{name}.db"
@ -1374,8 +1358,7 @@ async def test_write_thread_context_is_detached_between_tasks(
if threading.current_thread().name == write_thread_name:
depth["value"] -= 1
# Patched on the opentelemetry.context module itself, which is what both
# database.py and opentelemetry.trace.use_span() look the functions up on.
# database.py and opentelemetry.trace both call these via the module
monkeypatch.setattr(otel_context_api, "attach", counting_attach)
monkeypatch.setattr(otel_context_api, "detach", counting_detach)
@ -1388,8 +1371,6 @@ async def test_write_thread_context_is_detached_between_tasks(
def failing_probe(conn):
probe(conn)
# Exercises the write thread's exception path: the detach still has
# to happen, which is why it lives in a `finally`.
raise ValueError("deliberate failure inside a write task")
try:
@ -1405,9 +1386,7 @@ async def test_write_thread_context_is_detached_between_tasks(
finally:
real_detach(token)
# Sanity check: no marker is active in *this* (event loop) context
# right now, so the final probe is a fair test of the write thread's
# own state rather than something this test forgot to clean up.
# No marker is set here, so the final probe should see None
assert otel_context_api.get_value(_CONTEXT_LEAK_MARKER_KEY) is None
await db.execute_write_fn(probe)
finally:

329
tests/test_telemetry.py
View file

@ -26,8 +26,8 @@ SECRET_PARAM_VALUE = "SUPER_SECRET_PARAM_VALUE_XYZ_123"
INVALID_SQL = "select this_is_not_valid_sql from nowhere"
# Bounded so a broken time limit fails the test instead of hanging it, but far
# too long to finish inside any of the millisecond budgets used below.
# Bounded so a broken time limit fails rather than hangs, but too slow to
# finish within the millisecond time limits used below.
SLOW_SQL = """
with recursive counter(x) as (
select 1 union all select x + 1 from counter where x < 50000000
@ -41,13 +41,7 @@ def _db_query_spans(otel_spans):
def _spans_for_namespace(otel_spans, namespace):
"""
db.query spans belonging to one database.
Datasette queries its internal catalog constantly - including while a
Datasette instance is being constructed - so a test that just grabbed
every db.query span would be reading someone else's traffic.
"""
"db.query spans for one database, excluding queries against the internal database."
return [
span
for span in _db_query_spans(otel_spans)
@ -56,14 +50,7 @@ def _spans_for_namespace(otel_spans, namespace):
def _children_named(otel_spans, name, parent_span_context):
"""
Finished spans called `name` whose parent really is `parent_span_context`.
Parentage is matched on span id, not on "a span with this name exists" -
a span can exist and still be an unparented root if a thread boundary
dropped the otel context, which is the exact failure these tests exist
to catch.
"""
"Finished spans called `name` that are direct children of `parent_span_context`."
return [
span
for span in otel_spans.get_finished_spans()
@ -76,12 +63,7 @@ def _children_named(otel_spans, name, parent_span_context):
def _descends_from(span, ancestor_span_context, by_span_id):
"""
True if `span` reaches `ancestor_span_context` by walking parent links.
Walks real span ids rather than trusting a shared trace id: a span can
carry the right trace id and still hang off the wrong parent.
"""
"True if `span` reaches `ancestor_span_context` by walking parent links."
seen = set()
current = span
while current.parent is not None:
@ -97,7 +79,7 @@ def _descends_from(span, ancestor_span_context, by_span_id):
def _all_attribute_values(otel_spans):
"Every attribute value across every finished span, for the 'no leaked param values' test."
"Every attribute value on every finished span and span event."
values = []
for span in otel_spans.get_finished_spans():
values.extend((span.attributes or {}).values())
@ -108,14 +90,9 @@ def _all_attribute_values(otel_spans):
def test_datasette_package_never_imports_the_sdk():
"""
Core depends on opentelemetry-api only. The SDK is a test dependency.
Importing datasette does not load the OpenTelemetry SDK.
Checked by importing datasette in a fresh process and inspecting
sys.modules, rather than by grepping, so a lazy `import
opentelemetry.sdk` inside a function body cannot slip past.
conftest.py's pytest_collection_modifyitems() moves this test to the
front of the run by name - if you rename it, rename it there too.
conftest.py moves this test to the front of the run by name.
"""
code = (
"import datasette.app, datasette.database, datasette.telemetry, sys; "
@ -149,13 +126,7 @@ async def test_db_query_span_basic_attributes(ds_client, otel_spans):
@pytest.mark.asyncio
async def test_truncated_result_sets_truncated_attribute(otel_spans):
"""
A result actually cut short by max_returned_rows records truncated=True.
Every other test asserts the attribute is False, so a regression that
recorded the flag before the slice (or inverted it) would pass the rest
of the suite.
"""
"A result cut short by max_returned_rows records truncated=True."
ds = Datasette(memory=True, settings={"max_returned_rows": 5})
db = ds.add_memory_database("t04_truncated")
results = await db.execute(
@ -178,8 +149,7 @@ async def test_facetable_request_produces_db_query_spans(ds_client, otel_spans):
spans = _db_query_spans(otel_spans)
assert spans, "expected at least one db.query span"
assert all(span.attributes["db.system"] == "sqlite" for span in spans)
# Every db.query names what ran: SQL text for the string methods,
# datasette.callback for callback-style calls (schema introspection here).
# Each span records the SQL or, for callback methods, the callback name:
assert all(
span.attributes.get("db.query.text")
or span.attributes.get("datasette.callback")
@ -194,8 +164,7 @@ async def test_facetable_request_produces_db_query_spans(ds_client, otel_spans):
def test_sql_attribute_truncates_at_2048():
short_sql = "select 1"
assert sql_attribute(short_sql) == "select 1"
# Whitespace is stripped, so the same query logged twice with different
# surrounding whitespace produces one attribute value, not two.
# Surrounding whitespace is stripped:
assert sql_attribute(" select 1\n") == "select 1"
long_sql = "select 1 -- " + ("x" * 3000)
@ -207,8 +176,7 @@ def test_sql_attribute_truncates_at_2048():
@pytest.mark.asyncio
async def test_db_query_text_is_truncated_in_real_span(ds_client, otel_spans):
# A long trailing SQL comment keeps the query valid and executable while
# pushing db.query.text well past the 2048 char cap.
# A long trailing comment keeps the SQL valid but over the 2048 character limit
long_sql = "select 1 -- " + ("x" * 3000)
response = await ds_client.get("/fixtures/-/query.json", params={"sql": long_sql})
assert response.status_code == 200
@ -231,8 +199,7 @@ async def test_no_span_attribute_ever_contains_a_parameter_value(ds_client, otel
params={"sql": "select :secret", "secret": SECRET_PARAM_VALUE},
)
assert response.status_code == 200
# Sanity check the value really did flow through as a bound parameter,
# not inlined into the SQL text, otherwise this test would be vacuous.
# Confirm the bound parameter value was used by the query:
assert SECRET_PARAM_VALUE in json.dumps(response.json())
for value in _all_attribute_values(otel_spans):
@ -253,13 +220,10 @@ async def test_no_span_attribute_ever_contains_a_parameter_value(ds_client, otel
@pytest.mark.asyncio
async def test_query_interrupted_sets_error_status(otel_spans):
"""
A query that runs out the instance-wide sql_time_limit_ms is an error.
A query that exceeds the sql_time_limit_ms setting is a span error.
This used to force the timeout with `?_timelimit=5`, but a caller-supplied
budget shorter than the instance limit is now the signal that the timeout
was expected - see test_expected_timeout_is_not_a_span_error - so the
timeout has to come from the setting for this to still test what it was
written to test.
The limit comes from the setting because a shorter custom_time_limit
marks the timeout as expected.
"""
ds = Datasette(memory=True, settings={"sql_time_limit_ms": 20})
db = ds.add_memory_database("t09_instance_limit_timeout")
@ -276,23 +240,14 @@ async def test_query_interrupted_sets_error_status(otel_spans):
async def _expected_timeout_count_span(otel_spans, database_name):
"""
Drive the real table_counts() path into a timeout; return its db.query span.
table_counts() is where the headline instance of this lives: the homepage
counts every table under a 10ms budget and stores None for any table that
does not finish in time. Before this was fixed, a two-table database
produced four ERROR spans - two db.query and two db.query.execute - on
every single homepage hit.
"""
"Make table_counts() time out and return its db.query span."
db = Datasette(memory=True).add_memory_database(database_name)
await db.execute_write("create table big (id integer primary key, t text)")
await db.execute_write_many(
"insert into big (t) values (?)", [["x" * 50] for _ in range(11000)]
)
# count_limit caps the scan at 10001 rows, and below 20ms sqlite_timelimit()
# runs its progress handler on every VM instruction, so 1ms is not a close
# call - a scan of that size takes single-digit milliseconds at best.
# count_limit caps the scan at 10001 rows. Below 20ms sqlite_timelimit()
# checks the limit on every VM instruction, so this reliably exceeds 1ms.
counts = await db.table_counts(1)
assert counts == {
"big": None
@ -310,7 +265,7 @@ async def _expected_timeout_count_span(otel_spans, database_name):
@pytest.mark.asyncio
async def test_expected_timeout_is_not_a_span_error(otel_spans):
span = await _expected_timeout_count_span(otel_spans, "t09_expected_timeout")
# The useful signal survives; only the red status goes away.
# Recorded as interrupted, but not as an error:
assert span.attributes["datasette.interrupted"] is True
assert span.status.status_code != StatusCode.ERROR
assert not [event for event in span.events if event.name == "exception"]
@ -318,13 +273,7 @@ async def test_expected_timeout_is_not_a_span_error(otel_spans):
@pytest.mark.asyncio
async def test_expected_timeout_does_not_error_the_inner_execute_span(otel_spans):
"""
The same fix has to reach db.query.execute, which sets its own status.
Half of the original bug lived here: the inner span passed
set_status_on_exception=log_sql_errors, and table_counts() leaves
log_sql_errors at its True default, so it went ERROR too.
"""
"The db.query.execute child span is not marked as an error either."
span = await _expected_timeout_count_span(otel_spans, "t09_expected_timeout_inner")
children = _children_named(otel_spans, "db.query.execute", span.context)
assert len(children) == 1
@ -335,13 +284,7 @@ async def test_expected_timeout_does_not_error_the_inner_execute_span(otel_spans
@pytest.mark.asyncio
async def test_unexpected_timeout_is_still_a_span_error(otel_spans):
"""
A custom_time_limit *above* sql_time_limit_ms is not a short budget.
This is the half of the rule that stops the fix collapsing into "never
report timeouts": the caller asked for 5 seconds, the instance overruled it
at 20ms, and nobody expected that.
"""
"A timeout is an error if custom_time_limit is above sql_time_limit_ms."
ds = Datasette(memory=True, settings={"sql_time_limit_ms": 20})
db = ds.add_memory_database("t09_custom_limit_ignored")
with pytest.raises(QueryInterrupted):
@ -350,8 +293,7 @@ async def test_unexpected_timeout_is_still_a_span_error(otel_spans):
spans = _spans_for_namespace(otel_spans, "t09_custom_limit_ignored")
assert spans
span = spans[-1]
# Proves the caller's larger budget really was discarded - otherwise this
# would be asserting on a query that ran under a 5s limit.
# The setting overrides the larger custom_time_limit:
assert span.attributes["datasette.time_limit_ms"] == 20
assert span.attributes["datasette.interrupted"] is True
assert span.status.status_code == StatusCode.ERROR
@ -378,14 +320,7 @@ async def test_unsuppressed_sql_error_is_a_span_error(ds_client, otel_spans):
@pytest.mark.asyncio
async def test_suppressed_sql_error_is_not_a_span_error(ds_client, otel_spans):
"""
log_sql_errors=False means the caller is probing and expects failures.
Facet suggestion runs `json_type(column)` against every column precisely
to discover which ones raise, so marking those spans as errors would put
two red spans per text column on every table page - burying real failures
and tripping any alerting keyed on span status.
"""
"With log_sql_errors=False the error is recorded as suppressed, not a span error."
db = ds_client.ds.get_database("fixtures")
with pytest.raises(sqlite3.OperationalError):
await db.execute(INVALID_SQL, log_sql_errors=False)
@ -400,8 +335,7 @@ async def test_suppressed_sql_error_is_not_a_span_error(ds_client, otel_spans):
@pytest.mark.asyncio
async def test_execute_write_produces_db_query_span(otel_spans):
# Named in-memory databases are shared-cache, so every test in this file
# needs its own name or the second `create table` hits an existing table.
# Named in-memory databases are shared, so each test uses a unique name.
db = Datasette(memory=True).add_memory_database("t03_write_span")
await db.execute_write("create table docs (id integer primary key, name text)")
await db.execute_write("insert into docs (id, name) values (?, ?)", [1, "one"])
@ -451,27 +385,18 @@ async def test_execute_write_many_records_param_sets_not_rows_returned(otel_span
span = many_spans[0]
assert span.attributes["datasette.param_sets"] == 5
# executemany() consumes parameter sets and returns no rows at all, so
# calling this a row count would be a lie. Asserted explicitly because the
# attribute really was named datasette.rows_returned at one point.
assert "datasette.rows_returned" not in span.attributes
# --- Context propagation across thread boundaries --------------------------
#
# Every assertion below checks parentage (child.parent.span_id ==
# expected_parent.span_id, in the same trace), not merely that spans exist.
# Spans can exist and still be wrongly parented - or be unparented roots - if
# a thread boundary drops the otel context, which is exactly the failure mode
# these tests exist to prevent.
# These tests check span parentage, not just that the spans exist.
@pytest.mark.asyncio
async def test_db_query_execute_parents_to_db_query(ds_client, otel_spans):
# execute_fn()'s executor.submit() is thread boundary #1. The
# db.query.execute span is created inside the worker thread; without the
# copy_context() propagation it comes back as an unparented root span
# rather than a child of db.query.
# execute_fn() submits to the executor, so db.query.execute is created on
# another thread.
response = await ds_client.get("/fixtures/-/query.json?sql=select+1")
assert response.status_code == 200
@ -490,19 +415,15 @@ async def test_db_query_execute_parents_to_db_query(ds_client, otel_spans):
], "expected at least one db.query.execute span"
children = _children_named(otel_spans, "db.query.execute", query_span.context)
assert len(children) == 1, "expected exactly one db.query.execute child of db.query"
# The execute span is strictly contained by the round-trip span, and the
# gap between the two is the thread-pool wait.
# db.query.execute runs within db.query; the gap is the thread pool wait.
assert query_span.start_time <= children[0].start_time
assert children[0].end_time <= query_span.end_time
@pytest.mark.asyncio
async def test_immutable_database_propagates_context(tmp_path, otel_spans):
# Thread boundary #3, the easy one to miss: immutable databases route
# execute_isolated_fn() through loop.run_in_executor() directly rather
# than through the write thread. A span created inside that worker must
# still parent to whatever was current when execute_isolated_fn() was
# awaited, or every immutable-database operation emits orphan roots.
# Immutable databases run execute_isolated_fn() on another thread using
# loop.run_in_executor(), not the write thread.
db_path = tmp_path / "t04_immutable.db"
sqlite_utils.Database(str(db_path))["t"].insert({"id": 1}, pk="id")
@ -526,10 +447,7 @@ async def test_immutable_database_propagates_context(tmp_path, otel_spans):
for span in otel_spans.get_finished_spans()
if span.name == "t04-child-in-isolated-worker"
], "expected a span created inside execute_isolated_fn's worker thread"
# execute_isolated_fn() now opens its own db.query span, so the chain is
# event-loop parent -> db.query -> worker child. The worker child
# parenting to that db.query span, across the thread, is the propagation
# this test exists to prove.
# Expected chain: event loop parent -> db.query -> worker thread child
query_spans = _children_named(otel_spans, "db.query", parent_context)
assert len(query_spans) == 1
children = _children_named(
@ -540,10 +458,8 @@ async def test_immutable_database_propagates_context(tmp_path, otel_spans):
@pytest.mark.asyncio
async def test_write_spans_parent_to_db_query(otel_spans):
# Thread boundary #2: WriteTask -> queue.Queue -> the write thread.
# db.write.queue_wait and db.write.execute are both direct children of
# the db.query span that was current on the event loop at enqueue time,
# so they are siblings rather than nested inside one another.
# execute_write() queues a WriteTask for the write thread.
# db.write.queue_wait and db.write.execute are both children of db.query.
db = Datasette(memory=True).add_memory_database("t04_write_spans")
await db.execute_write("create table docs (id integer primary key)")
@ -563,18 +479,14 @@ async def test_write_spans_parent_to_db_query(otel_spans):
execute_span = execute_children[0]
assert execute_span.attributes["datasette.isolated_connection"] is False
assert execute_span.attributes["datasette.transaction"] is True
# Siblings, not parent/child: the queue wait is over by the time the
# write begins.
# The queue wait ends before the write begins.
assert queue_wait_children[0].end_time <= execute_span.start_time
@pytest.mark.asyncio
async def test_write_queue_wait_duration_reflects_real_wait(otel_spans):
# db.write.queue_wait is built from explicit start/end timestamps -
# task.enqueued_at_ns, captured on the event loop, through to the moment
# the write thread dequeued it. If it were a plain `with` block on the
# write thread it would instead measure the microseconds spent building
# the span object, and this assertion would fail.
# db.write.queue_wait runs from task.enqueued_at_ns, captured on the event
# loop, to when the write thread dequeues the task.
ds = Datasette(memory=True)
db = ds.add_memory_database("t04_queue_wait")
await db.execute_write("create table docs (id integer primary key)")
@ -582,9 +494,7 @@ async def test_write_queue_wait_duration_reflects_real_wait(otel_spans):
def slow_write(conn):
time.sleep(0.1)
# Queue a deliberately slow write without waiting for it, then queue a
# second write immediately behind it: the second task sits in the queue
# for roughly the duration of the first.
# Queue a slow write without waiting for it, then a second write behind it:
_, slow_future = await db._send_to_write_thread(slow_write, block=False)
await db.execute_write("insert into docs (id) values (1)")
await slow_future
@ -600,23 +510,17 @@ async def test_write_queue_wait_duration_reflects_real_wait(otel_spans):
)
assert len(queue_wait_children) == 1
duration_ns = queue_wait_children[0].end_time - queue_wait_children[0].start_time
# The slow write sleeps 100ms; anything above 10ms is far beyond the
# microseconds a mis-timestamped span would report.
# The slow write sleeps for 100ms
assert duration_ns > 10_000_000, f"queue wait was only {duration_ns}ns"
async def _write_spans_from_one_enqueue(otel_spans, name, block):
"""
Run exactly one write through the write thread from inside a span of our
own, and return (enqueueing span context, {span name: span}).
Run one write through the write thread inside a span, returning
(enqueueing span context, {span name: span}).
`_send_to_write_thread` is called directly rather than `execute_write()`
because `execute_write()` opens its own db.query span, which would then
be the span current at enqueue time - so the parent/link would point at
that span rather than at the one this test controls.
The exporter is cleared immediately before the enqueue so the write spans
collected here can only have come from this one write.
Uses _send_to_write_thread() because execute_write() would add its own
db.query span between the enqueueing span and the write spans.
"""
db = Datasette(memory=True).add_memory_database(name)
await db.execute_write("create table docs (id integer primary key)")
@ -629,11 +533,8 @@ async def _write_spans_from_one_enqueue(otel_spans, name, block):
enqueuer_context = enqueuer.get_span_context()
queued = await db._send_to_write_thread(insert, block=block)
if not block:
# The point of block=False is that the write happens after the
# caller has returned and the enqueueing span above has closed.
# Awaiting the reply future outside that `with` waits for the write
# thread deterministically - it is resolved only after both write
# spans have ended and been exported.
# Wait for the write after the enqueueing span has ended. The reply
# future resolves once both write spans have been exported.
_, reply_future = queued
await reply_future
@ -648,9 +549,8 @@ async def _write_spans_from_one_enqueue(otel_spans, name, block):
@pytest.mark.asyncio
async def test_blocking_write_spans_still_parent_normally(otel_spans):
# Regression guard for ticket 07: block=True genuinely has containment -
# the caller awaits the reply future - so those spans must keep parenting
# to the enqueueing span, and must not grow links.
# block=True waits for the write, so its spans are children of the
# enqueueing span, with no links.
enqueuer_context, spans = await _write_spans_from_one_enqueue(
otel_spans, "t07_blocking_write", block=True
)
@ -664,24 +564,21 @@ async def test_blocking_write_spans_still_parent_normally(otel_spans):
@pytest.mark.asyncio
async def test_nonblocking_write_spans_are_roots_with_a_link(otel_spans):
# block=False returns before the write runs, so the enqueueing span has
# already ended (and exported) by the time these spans start. Parenting
# them to it would draw a child outliving its closed parent, so they are
# roots in their own traces, linked back to the span that caused them.
# block=False returns before the write runs, so the write spans are roots
# linked to the enqueueing span.
enqueuer_context, spans = await _write_spans_from_one_enqueue(
otel_spans, "t07_nonblocking_write", block=False
)
assert enqueuer_context.is_valid, "test's own enqueueing span was not recorded"
for name, span in spans.items():
assert span.parent is None, f"{name} is still parented"
# A link does not join the linked trace: each of these is its own
# root trace, which is the correct shape and not a workaround.
# Each write span starts its own trace
assert span.context.trace_id != enqueuer_context.trace_id, name
assert len(span.links) == 1, f"{name} has links {span.links}"
link_context = span.links[0].context
assert link_context.trace_id == enqueuer_context.trace_id, name
assert link_context.span_id == enqueuer_context.span_id, name
# The two write spans are independent roots, not nested in one another.
# The two write spans are separate roots
assert (
spans["db.write.queue_wait"].context.trace_id
!= spans["db.write.execute"].context.trace_id
@ -690,8 +587,6 @@ async def test_nonblocking_write_spans_are_roots_with_a_link(otel_spans):
@pytest.mark.asyncio
async def test_nonblocking_write_link_has_no_attributes(otel_spans):
# There is only one kind of link here, so a relationship-name attribute
# would be a constant conveying nothing the link's existence does not.
_, spans = await _write_spans_from_one_enqueue(
otel_spans, "t07_nonblocking_link_attrs", block=False
)
@ -705,18 +600,10 @@ async def test_nonblocking_write_spans_ignore_the_write_threads_ambient_context(
otel_spans,
):
"""
block=False spans pass an explicit empty Context, not merely "no attach".
block=False spans ignore any context left attached on the write thread.
Nothing is attached for a block=False task, but "nothing attached" is not
the same as "no ambient context": the write thread is persistent, and
anything running on it - a prepare_connection plugin hook, say - can
attach a context and never detach it. Without the explicit `context=`
these spans would silently parent to that leftover span instead of being
roots, and no other test here would notice, because in every other test
the write thread's ambient context happens to be empty.
So this test leaks exactly such a context on the write thread, the way a
careless plugin would, and then checks the write spans are still roots.
A prepare_connection hook could attach a context and never detach it.
This test does that, then checks the write spans are still roots.
"""
ds = Datasette(memory=True)
db = ds.add_memory_database("t07_ambient_write_thread")
@ -726,8 +613,8 @@ async def test_nonblocking_write_spans_ignore_the_write_threads_ambient_context(
def prepare_connection(conn, database):
if threading.current_thread().name == write_thread_name:
# Runs once, on the write thread, before any task is dequeued -
# and never detaches, which is the whole point.
# Runs on the write thread before any task is dequeued, and never
# detaches.
span = tracer.start_span("leaked-write-thread-ambient-span")
leaked["span_id"] = span.get_span_context().span_id
otel_context_api.attach(otel_trace.set_span_in_context(span))
@ -766,14 +653,7 @@ async def test_nonblocking_write_spans_ignore_the_write_threads_ambient_context(
@pytest.mark.asyncio
async def test_suppressed_error_does_not_mark_execute_span(ds_client, otel_spans):
"""
The inner db.query.execute span must honour log_sql_errors too.
It is created inside the worker thread, so without record_exception /
set_status_on_exception being passed through it would mark every facet
suggestion probe as failed even though the outer db.query span correctly
reports the failure as suppressed.
"""
"The inner db.query.execute span also respects log_sql_errors=False."
db = ds_client.ds.get_database("fixtures")
with pytest.raises(sqlite3.OperationalError):
await db.execute(INVALID_SQL, log_sql_errors=False)
@ -791,23 +671,14 @@ async def test_suppressed_error_does_not_mark_execute_span(ds_client, otel_spans
@pytest.mark.asyncio
async def test_invoke_startup_produces_one_trace_not_dozens_of_orphans(otel_spans):
"""
invoke_startup() runs with no request, so nothing it does has an ambient
span to nest under. Without datasette.startup every register_* hook, every
internal-catalog read and every catalog write becomes its own single-span
root trace - around twenty of them per fresh instance.
"""
"Spans emitted by invoke_startup() share a single datasette.startup root span."
ds = Datasette(memory=True)
# Named in-memory databases are shared-cache, so this needs its own name.
ds.add_memory_database("t05_startup_db")
# Constructing a Datasette already touches the internal catalog, and that
# work is genuinely outside startup. Clear so the assertions below describe
# invoke_startup() alone.
# Ignore spans from constructing Datasette, which happens before startup
otel_spans.clear()
# Deliberately no ambient span: this mirrors the ASGI lifespan path, where
# startup runs before any request exists. If something did wrap this call
# the "one root" assertion below would pass for the wrong reason.
# No ambient span, as in the ASGI lifespan path where startup runs before
# any request.
assert (
not otel_trace.get_current_span().get_span_context().is_valid
), "this test must run with no ambient span"
@ -833,7 +704,7 @@ async def test_invoke_startup_produces_one_trace_not_dozens_of_orphans(otel_span
by_span_id = {span.context.span_id: span for span in spans}
# The internal catalog reads are what made up the bulk of the orphans.
# Internal database reads:
internal_queries = [
span
for span in spans
@ -844,8 +715,7 @@ async def test_invoke_startup_produces_one_trace_not_dozens_of_orphans(otel_span
_descends_from(span, startup.context, by_span_id) for span in internal_queries
)
# ...and the catalog writes, which reach the span through the write thread,
# so they also prove the ticket-04 context capture survives startup.
# Internal database writes, which run on the write thread:
write_spans = [span for span in spans if span.name.startswith("db.write.")]
assert write_spans, "expected db.write.* spans during startup"
assert all(
@ -859,20 +729,11 @@ async def test_invoke_startup_produces_one_trace_not_dozens_of_orphans(otel_span
@pytest.mark.asyncio
async def test_db_query_is_client_kind_and_children_are_internal(otel_spans):
"""
db.query is a database client span; Datasette's decomposition of it is not.
Trace UIs key their database rendering off the span kind rather than off
db.system, so db.query has to be CLIENT. db.query.execute,
db.write.execute and db.write.queue_wait deliberately stay INTERNAL: they
are parts of one logical query rather than three separate database calls,
and queue_wait touches no database at all - marking them CLIENT would
make one query look like several to anything counting spans by kind.
db.query spans are CLIENT. Their child spans are INTERNAL because they are
parts of one query rather than separate database calls.
"""
# Named in-memory databases are shared-cache, so this needs its own name.
db = Datasette(memory=True).add_memory_database("t06_span_kind")
# All four db.query entry points, so a missed `kind=` on any one of them
# fails here - plus the write path (db.write.queue_wait,
# db.write.execute) and the read path (db.query.execute) children.
# Call each of the four SQL string methods:
await db.execute_write("create table docs (id integer primary key)")
await db.execute_write_many(
"insert into docs (id) values (?)", [[i] for i in range(1, 4)]
@ -899,15 +760,7 @@ async def test_db_query_is_client_kind_and_children_are_internal(otel_spans):
async def test_instrumentation_scope_declares_version_and_schema_url(
ds_client, otel_spans
):
"""
Spans say which Datasette produced them and which semconv version their
attribute names follow.
Before get_tracer() was given a version and a schema URL every exported
scope was name='datasette' version='' schema_url='', so nothing
downstream could tell which Datasette a span came from, or whether
`db.system` meant `db.system` or the post-1.30.0 `db.system.name`.
"""
"The instrumentation scope includes the Datasette version and schema URL."
response = await ds_client.get("/fixtures/-/query.json?sql=select+1")
assert response.status_code == 200
@ -917,10 +770,7 @@ async def test_instrumentation_scope_declares_version_and_schema_url(
assert scope.name == "datasette"
assert scope.version == __version__
# The literal URL, not the SCHEMA_URL constant: comparing the span
# against the same constant the instrumentation is built from would only
# catch a dropped argument, never a wrong value. Bumping this is a claim
# about the attribute names on the wire - see SCHEMA_URL in telemetry.py.
# Uses the literal URL so changing SCHEMA_URL requires updating this test
assert scope.schema_url == "https://opentelemetry.io/schemas/1.29.0"
assert SCHEMA_URL == "https://opentelemetry.io/schemas/1.29.0"
assert __version__, "the scope version must not be empty"
@ -929,14 +779,11 @@ async def test_instrumentation_scope_declares_version_and_schema_url(
def test_db_operation_name_from_leading_keyword():
assert sql_operation_name("select 1") == "SELECT"
assert sql_operation_name(" insert into x (a) values (1)") == "INSERT"
# A leading CTE reports WITH rather than the operation inside it. That is
# the documented limitation, not an accident - see sql_operation_name().
# A leading CTE reports WITH, not the operation inside it
assert sql_operation_name("with foo as (select 1) select * from foo") == "WITH"
# Unrecognised leading keyword: no attribute rather than a wrong one, and
# no unbounded value set derived from attacker-supplied SQL.
# Unrecognized leading keyword
assert sql_operation_name("gibberish 1") is None
# Not a parser: a parenthesised SELECT and a leading comment both yield
# nothing rather than a guess.
# A parenthesized SELECT or a leading comment also returns None
assert sql_operation_name("(select 1) union select 2") is None
assert sql_operation_name("-- a comment\nselect 1") is None
assert sql_operation_name("") is None
@ -976,14 +823,10 @@ async def test_execute_write_sets_db_operation_name(otel_spans):
@pytest.mark.asyncio
async def test_execute_write_script_has_no_operation_name(otel_spans):
"""
executescript() runs several statements, so naming the operation after
the first one would be a lie. Semantic conventions say db.operation.name
should not be extracted from query text that can hold more than one
operation, so the attribute is absent entirely.
Scripts can contain several statements, so db.operation.name is omitted.
The script deliberately starts with `create`, which *is* on the
allowlist - so this fails if the call site ever starts calling
sql_operation_name().
The script starts with `create`, which is on the allowlist, so this fails
if the operation name is extracted anyway.
"""
db = Datasette(memory=True).add_memory_database("t06_script_operation")
await db.execute_write_script(
@ -1022,8 +865,7 @@ async def test_execute_fn_produces_db_query_span(otel_spans):
span.attributes["datasette.callback"]
== "test_execute_fn_produces_db_query_span.<locals>.count_rows"
)
# There is no SQL string for a callback, and no statement to take a
# leading keyword from - absent beats guessed.
# Callbacks have no SQL text to record or take an operation name from
assert "db.query.text" not in span.attributes
assert "db.operation.name" not in span.attributes
children = _children_named(otel_spans, "db.query.execute", span.context)
@ -1032,7 +874,6 @@ async def test_execute_fn_produces_db_query_span(otel_spans):
@pytest.mark.asyncio
async def test_execute_fn_lambda_reports_lambda(otel_spans):
# Pins the documented behaviour rather than pretending lambdas have names.
db = Datasette(memory=True).add_memory_database("t16_lambda")
otel_spans.clear()
await db.execute_fn(lambda conn: conn.execute("select 1").fetchone())
@ -1067,9 +908,7 @@ async def test_execute_write_fn_produces_db_query_span(otel_spans):
@pytest.mark.asyncio
async def test_execute_write_fn_callback_name_is_not_the_hook_wrapper(otel_spans):
# A callback that declares track_event is the case where
# _wrap_fn_with_hooks() actually replaces fn with a wrapper - the span
# must still report the caller's function, not the wrapper's name.
# _wrap_fn_with_hooks() wraps callbacks that accept track_event
db = Datasette(memory=True).add_memory_database("t16_wrapper_name")
def create_with_events(conn, track_event):
@ -1087,9 +926,8 @@ async def test_execute_write_fn_callback_name_is_not_the_hook_wrapper(otel_spans
@pytest.mark.asyncio
async def test_execute_write_fn_nonblocking_spans_link_to_the_new_span(otel_spans):
# For block=False the public db.query span ends at enqueue and the
# write-thread spans become roots. Their link must target that new span,
# not whatever was current around the execute_write_fn() call.
# With block=False the write thread spans link to the db.query span from
# execute_write_fn(), not to the span that was current when it was called.
db = Datasette(memory=True).add_memory_database("t16_nonblocking")
await db.execute_write("create table docs (id integer primary key)")
@ -1100,8 +938,7 @@ async def test_execute_write_fn_nonblocking_spans_link_to_the_new_span(otel_span
with tracer.start_as_current_span("t16-enqueueing-span") as enqueuer:
enqueuer_context = enqueuer.get_span_context()
await db.execute_write_fn(insert, block=False)
# Writes are serialized on the write thread, so a blocking write behind
# the non-blocking one waits for it deterministically.
# Writes run in order, so this waits for the non-blocking write to finish
await db.execute_write("insert into docs (id) values (2)")
query_spans = [
@ -1125,9 +962,8 @@ async def test_execute_write_fn_nonblocking_spans_link_to_the_new_span(otel_span
@pytest.mark.asyncio
async def test_execute_does_not_double_wrap(otel_spans):
# The regression guard for the refactor: execute() and the SQL-string
# write methods call the private _execute_fn/_execute_write_fn, so they
# must not gain a second db.query span from the public wrappers.
# execute() and the SQL string write methods call the private
# _execute_fn() and _execute_write_fn(), so they create one db.query span.
db = Datasette(memory=True).add_memory_database("t16_no_double_wrap")
otel_spans.clear()
await db.execute_write("create table t (id integer primary key)")
@ -1172,8 +1008,7 @@ async def test_execute_isolated_fn_span_on_mutable_and_immutable(tmp_path, otel_
@pytest.mark.asyncio
async def test_execute_fn_exception_marks_span_error(otel_spans):
# Unlike execute(), there is no probing caller on this path - a callback
# that raises is an error, with the default record_exception behaviour.
# execute_fn() has no log_sql_errors option, so exceptions are span errors
db = Datasette(memory=True).add_memory_database("t16_fn_error")
def boom(conn):

95
tests/test_telemetry_metrics.py
View file

@ -1,19 +1,6 @@
"""
Tests for the OpenTelemetry metrics Datasette core emits.
Two layers are tested separately and deliberately:
- The gauge callbacks are plain generator functions, so they are called
directly for exact-value assertions. Going through the SDK for those would
be unreliable: the pool gauges carry no attribute identifying which
Datasette produced them, and a pytest session has many live instances, so
the SDK's last-value aggregation would report whichever one happened to be
observed last.
- The SDK pipeline (instrument -> reader -> data points) is tested through
the `otel_metrics` fixture, using metrics that carry `db.namespace` - a
uniquely named in-memory database is enough to isolate those from every
other instance alive in the session.
Tests for the OpenTelemetry metrics emitted by Datasette. Gauge callbacks are
called directly, since the pool gauges have no attributes to tell instances apart.
"""
import asyncio
@ -65,25 +52,17 @@ def metrics_ds():
@pytest.mark.asyncio
async def test_sql_thread_limit_gauge_reports_num_sql_threads(metrics_ds):
values = [value for _, value in observations(telemetry.observe_sql_thread_limit)]
# Other instances are alive in this session, so assert membership rather
# than uniqueness - 7 is distinctive enough to only come from metrics_ds.
# Other Datasette instances may also be reporting:
assert 7 in values
@pytest.mark.asyncio
async def test_no_thread_gauges_in_non_threaded_mode():
"""
num_sql_threads=0 means there is no pool at all, so the pool gauges must
skip the instance rather than report a bogus limit of 0.
The pool gauges carry no attributes, so the live-instance registry is
narrowed to just this instance for the assertion - counting global
observations instead would let an unrelated instance being garbage
collected mid-test shift the baseline.
"""
"Pool gauges skip instances with num_sql_threads=0, which have no pool."
ds = Datasette(memory=True, settings={"num_sql_threads": 0})
try:
assert ds.executor is None
# Pool gauges have no attributes, so observe only this instance:
original = telemetry._live_datasettes
telemetry._live_datasettes = weakref.WeakSet([ds])
try:
@ -91,7 +70,7 @@ async def test_no_thread_gauges_in_non_threaded_mode():
assert list(telemetry.observe_sql_thread_queue_depth()) == []
finally:
telemetry._live_datasettes = original
# Per-database gauges are unaffected - they do not depend on the pool.
# Per-database gauges do not depend on the pool:
assert observations(telemetry.observe_pending_queries, ds)
finally:
ds.close()
@ -100,11 +79,8 @@ async def test_no_thread_gauges_in_non_threaded_mode():
@pytest.mark.asyncio
async def test_thread_queue_depth_gauge_reports_saturation():
"""
The headline alerting metric must actually read above zero when reads
queue behind num_sql_threads. This also pins the private
`ThreadPoolExecutor._work_queue` attribute the callback depends on: if a
stdlib rename ever removes it, this fails instead of the metric silently
vanishing (the callback tolerates its absence at collection time).
Queue depth is above zero when reads queue behind num_sql_threads. Also
fails if the private ThreadPoolExecutor._work_queue attribute goes away.
"""
ds = Datasette(memory=True, settings={"num_sql_threads": 1})
db = ds.add_memory_database("metrics_saturation_db")
@ -118,11 +94,10 @@ async def test_thread_queue_depth_gauge_reports_saturation():
try:
first = asyncio.ensure_future(db.execute_fn(blocker))
# Wait until the blocker owns the pool's only thread.
# Wait until the blocker is using the only thread:
await asyncio.get_running_loop().run_in_executor(None, entered.wait, 10)
second = asyncio.ensure_future(db.execute_fn(lambda conn: 2))
# The second submission lands in the executor's queue on the next
# event-loop turn; poll briefly rather than assume the timing.
# The second query is queued on a later event loop turn, so poll:
depths = []
for _ in range(500):
depths = [
@ -157,8 +132,7 @@ async def test_pending_queries_gauge_tracks_in_flight_queries(metrics_ds):
assert value() == 0
# sqlite3.sleep is not a thing, so block the worker thread on an event we
# control from the event loop and sample the gauge while it is held.
# Hold the worker thread until release is set:
release = asyncio.Event()
loop = asyncio.get_running_loop()
entered = asyncio.Event()
@ -188,8 +162,7 @@ async def test_write_queue_depth_gauge(metrics_ds):
if a == attributes
]
# No write has ever been queued, so there is no queue and no observation -
# rather than a fabricated zero for a queue that does not exist.
# No observation until the write queue has been created:
assert depths() == []
await db.execute_write("create table t (id integer primary key)")
@ -277,11 +250,7 @@ async def test_operation_duration_records_error_type(otel_metrics):
@pytest.mark.asyncio
async def test_operation_duration_records_write_error_type(otel_metrics):
"""
Same as the read-path error test, but the write wrappers time a different
code path - `execute_write_fn`, the write thread and its reply future -
so error propagation through them is pinned separately.
"""
"A failed write is still timed and records error.type."
ds = Datasette(memory=True)
ds.add_memory_database("duration_write_error_db")
try:
@ -319,7 +288,7 @@ async def test_write_queue_wait_histogram(otel_metrics):
@pytest.mark.asyncio
async def test_interrupted_queries_counter(otel_metrics):
"The count of time-limit kills, which sampled traces cannot provide."
"Queries cancelled by sql_time_limit_ms are counted."
ds = Datasette(memory=True, settings={"sql_time_limit_ms": 1})
ds.add_memory_database("interrupted_db")
try:
@ -344,7 +313,7 @@ async def test_interrupted_queries_counter(otel_metrics):
@pytest.mark.asyncio
async def test_metrics_are_reported_through_the_sdk_for_gauges(otel_metrics):
"End-to-end: a gauge callback reaches the reader as a data point."
"Gauge callbacks reach the metric reader as data points."
ds = Datasette(memory=True)
ds.add_memory_database("gauge_pipeline_db")
try:
@ -373,14 +342,8 @@ def test_closed_datasette_stops_being_observed():
def test_registry_holds_instances_weakly():
"""
Registering an instance must never be the thing that keeps it alive.
A stand-in object is used rather than a real Datasette because a Datasette
with a temp-disk internal database is pinned for the life of the process
by `Database.__init__`'s `atexit.register(self._cleanup_temp_file)`, which
holds the Database, which holds the Datasette. That is pre-existing and
unrelated to telemetry; what is tested here is that this registry adds no
reference of its own.
Registering an instance does not keep it alive. Uses a stand-in object
because an atexit handler in Database.__init__ keeps a real Datasette alive.
"""
import gc
import weakref
@ -412,10 +375,8 @@ HISTOGRAM_PROBES = [
),
]
# One value inside each of six distinct registry buckets. Under OpenTelemetry's
# default boundaries - [0, 5, 10, 25, ...], meant for milliseconds - the first
# five of these all land in (0, 5] and only 7.0 lands elsewhere, so the
# "occupies six buckets" assertion below fails if the advisory is ever dropped.
# One value in each of six registry buckets. The SDK's default boundaries
# would put the first five in the same bucket.
SPREAD = [0.00005, 0.0003, 0.002, 0.03, 0.8, 7.0]
@ -428,18 +389,8 @@ def test_histograms_spread_values_across_buckets(
otel_metrics, instrument_name, metric_name, attributes
):
"""
The registry's boundaries reach the SDK, and a realistic spread of
seconds-scale durations occupies more than one bucket.
Recording onto the instrument directly rather than driving a workload is
deliberate: real durations here are all tens of microseconds and would
share a bucket no matter what the boundaries were, which is exactly the
situation this test exists to detect.
`explicit_bounds` is compared against the registry rather than against the
instrument's own configuration - the instrument is built *from* the
registry, so that comparison would be a value against itself. What is
checked here is that the advisory survived the trip through the SDK.
The registry's bucket boundaries reach the SDK. Values are recorded
directly because real test query durations would all share one bucket.
"""
from datasette.telemetry_registry import METRICS
@ -464,7 +415,7 @@ def test_histograms_spread_values_across_buckets(
@pytest.mark.asyncio
async def test_operation_duration_histogram_records_execute_fn(otel_metrics):
"Callback-style reads land in the same histogram as SQL-string reads."
"execute_fn() reads are recorded in the same histogram as SQL reads."
ds = Datasette(memory=True)
ds.add_memory_database("duration_fn_db")
try:
@ -487,7 +438,7 @@ async def test_operation_duration_histogram_records_execute_fn(otel_metrics):
@pytest.mark.asyncio
async def test_operation_duration_histogram_records_execute_write_fn(otel_metrics):
"Callback-style writes - the JSON write API's whole diet - are counted too."
"execute_write_fn() writes are recorded in the same histogram."
ds = Datasette(memory=True)
ds.add_memory_database("duration_write_fn_db")
try:

242
tests/test_telemetry_registry.py
View file

@ -1,26 +1,6 @@
"""
Two-way conformance between `datasette/telemetry_registry.py` and what
Datasette actually emits.
This is the test that makes the generated documentation trustworthy. cog
guarantees the docs match the registry; this guarantees the registry matches
the code. Without it, both could agree with each other and be wrong.
It checks both directions, and the second one is the one nothing else catches:
- **emitted but not registered** - instrumentation was added without
documenting it, so the reference page silently omits it.
- **registered but never emitted** - the reference page describes a span or
attribute that no longer exists, which is worse than omitting it, because a
reader will build a dashboard on it.
Both of those directions compare the code against the registry. Neither can
catch a *rename*, because the call sites now take their names from the
registry - move `DB_NAMESPACE` to `"db.namespace2"` and code and registry
still agree with each other, while every existing dashboard breaks. So the
literal names live here too, spelled out, and are asserted against both the
registry and the wire. That is the one comparison in this file that is not
made against a value derived from the registry itself.
Tests that the spans, attributes and metrics Datasette emits match
datasette/telemetry_registry.py, in both directions.
"""
import copy
@ -42,10 +22,8 @@ from datasette.database import QueryInterrupted
from datasette.telemetry_testing import assert_metrics_conform, assert_metrics_covered
from datasette.utils.sqlite import sqlite3
# The names as they appear on the wire, written out rather than read from the
# registry. If a change to the registry makes one of these fail, that change
# is renaming something a user's dashboards and saved queries depend on -
# which is a decision to take deliberately, here, not a line to re-derive.
# Written out as literals rather than read from the registry, so renaming a
# signal fails these tests.
EXPECTED_ATTRIBUTES = {
"db.query": {
"db.system",
@ -73,18 +51,8 @@ EXPECTED_ATTRIBUTES = {
}
EXPECTED_SPANS = set(EXPECTED_ATTRIBUTES)
# The HTTP request span is handled separately because its name is composed at
# runtime - the request method, then the route it matched - so there is no
# fixed string to pin it to. What can still be pinned, and is what a dashboard
# depends on, is the shape of that name and the attribute keys.
#
# The route half is deliberately not spelled out as a literal: it is a core
# route regex, and pinning those here would make an unrelated routing change
# fail the telemetry conformance test. What is pinned instead is that the name
# is exactly the method, a space, and the span's own `http.route` value - the
# `{method} {route}` shape semantic conventions specify. The workload below
# only issues GETs, so a change that stopped clamping the method, or that
# started naming the span after the path, fails here.
# The HTTP request span name is composed at runtime as "{method} {route}", so
# it is checked by shape rather than as a literal. The workload only issues GETs.
EXPECTED_HTTP_SPAN_NAME = "{http.request.method} {http.route}"
EXPECTED_HTTP_METHOD_NAMES = {"GET"}
EXPECTED_HTTP_ATTRIBUTES = {
@ -99,16 +67,14 @@ EXPECTED_HTTP_ATTRIBUTES = {
"datasette.internal_client",
}
# The registry's own name for the request span is that template, not anything
# that appears on the wire.
# The registry uses the name template for the request span.
EXPECTED_REGISTRY_ATTRIBUTES = dict(
EXPECTED_ATTRIBUTES, **{EXPECTED_HTTP_SPAN_NAME: EXPECTED_HTTP_ATTRIBUTES}
)
EXPECTED_REGISTRY_NAMES = set(EXPECTED_REGISTRY_ATTRIBUTES)
# Named in-memory databases are shared-cache, so two Datasette instances using
# the same name share one SQLite database - and the second `create table`
# fails. Every workload below therefore gets its own name.
# Named in-memory databases are shared between instances, so each workload
# uses a unique name.
_names = itertools.count()
@ -117,14 +83,7 @@ def _unique(prefix):
class _BoomPlugin:
"""
A route that raises.
`error.type` on the request span is only ever set by a 5xx, and nothing
in Datasette returns one on a healthy instance - `route_path` converts
exceptions into a 500 itself, so the workload has to supply the
exception.
"""
"A route that raises, producing a 500 and error.type on the request span."
__name__ = "TelemetryRegistryBoomPlugin"
@ -135,20 +94,13 @@ class _BoomPlugin:
async def exercise():
"""
Drive enough of Datasette to emit every span and attribute the registry
claims exists.
Each call is here because it is the only thing that produces some span or
attribute - see the comments. If you add instrumentation on a path this
does not reach, add the path rather than loosening the assertions.
Returns the instance so the caller can close it; startup happens inside
so that the `datasette.startup` span lands in the collected set.
Drive enough of Datasette to emit every registered span and attribute,
including datasette.startup. Returns the instance so the caller can close it.
"""
name = _unique("registry")
ds = Datasette(memory=True)
ds.add_memory_database(name)
# datasette.startup - and the internal catalog work nested under it
# datasette.startup
await ds.invoke_startup()
db = ds.get_database(name)
@ -165,8 +117,7 @@ async def exercise():
# datasette.isolated_connection=True
await db.execute_isolated_fn(lambda conn: conn.execute("select 1").fetchone())
# datasette.callback, with named functions so the conformance run sees the
# attribute's documented value shape (a qualname, not just "<lambda>")
# datasette.callback, using named functions rather than lambdas
def registry_read_callback(conn):
return conn.execute("select count(*) from t").fetchone()
@ -181,14 +132,11 @@ async def exercise():
await db.execute("select * from t where id > :n", {"n": 5})
await db.execute("select * from t", truncate=True)
# datasette.sql_error_suppressed - the caller is probing and treats
# failure as an expected answer
# datasette.sql_error_suppressed
with pytest.raises(sqlite3.OperationalError):
await db.execute("select nope from t", log_sql_errors=False)
# datasette.interrupted - only ever set when a query exceeds its time
# limit, so the workload has to force one rather than exempt it. An
# unbounded recursive CTE cannot finish, so 1ms is always exceeded.
# datasette.interrupted: an unbounded recursive CTE always exceeds 1ms
with pytest.raises(QueryInterrupted):
await db.execute(
"with recursive c(x) as (select 0 union all select x+1 from c) "
@ -196,13 +144,11 @@ async def exercise():
custom_time_limit=1,
)
# These requests produce the HTTP request span and its
# http.request.method / url.path / url.scheme / server.address /
# user_agent.original / http.response.status_code attributes.
# HTTP request spans and their attributes
assert (await ds.client.get(f"/{name}/t?_facet=v")).status_code == 200
assert (await ds.client.get(f"/{name}/t/1.json")).status_code == 200
# error.type on the request span, which only a 5xx sets
# error.type on the request span, set by a 5xx response
ds.pm.register(_BoomPlugin(), name="telemetry-registry-boom")
try:
response = await ds.client.get("/-/telemetry-registry-boom")
@ -215,21 +161,13 @@ async def exercise():
@pytest_asyncio.fixture
async def emitted(otel_spans):
"""
Every (span name, span kind, attributes) triple a broad workload emits.
The kind is carried because the request span's name is composed at
runtime, so `span_for()` resolves it by kind instead. The attributes are
carried as a mapping rather than a set of keys because the request span's
name has to be checked against its own `http.route` value.
Every (span name, span kind, attributes) triple emitted by exercise().
The kind is needed to resolve the dynamically named request span.
"""
# otel_spans has already cleared the exporter, and nothing is cleared
# after this point: the workload's own startup emits datasette.startup.
ds = await exercise()
spans = otel_spans.get_finished_spans()
assert spans, "no spans captured - the fixture is not exercising anything"
# str() because span.name is the registry's SpanName instance, and a set
# of those would compare equal to literals but read confusingly in a
# failure message.
# str() so failure messages show plain strings, not registry instances
collected = tuple(
(
str(span.name),
@ -258,12 +196,7 @@ def _keys_by_span(records):
@pytest.mark.asyncio
async def test_workload_emits_exactly_the_expected_names(emitted):
"""
The wire format, pinned to literals.
Not derived from the registry, so this is what catches a rename that the
registry and the call sites make together.
"""
"Emitted span and attribute names match the expected literals."
static, server = _partition(emitted)
by_span = _keys_by_span(static)
assert set(by_span) == EXPECTED_SPANS
@ -275,9 +208,7 @@ async def test_workload_emits_exactly_the_expected_names(emitted):
for name, _kind, attributes in server:
union |= set(attributes)
route = attributes.get("http.route")
# Every request in the workload matches a route, so every one of these
# names must be `{method} {route}`. A 404 would be a bare method - the
# http_route tests cover that case with a real request.
# Every request in the workload matches a route
assert route, f"the request span {name!r} carries no http.route"
method, _, name_route = name.partition(" ")
assert name_route == route, (
@ -290,7 +221,7 @@ async def test_workload_emits_exactly_the_expected_names(emitted):
def test_registry_matches_the_expected_names():
"The other half of the rename check: the registry against the same literals."
"Registry names match the expected literals."
assert {str(span) for span in reg.SPANS} == EXPECTED_REGISTRY_NAMES
for span in reg.SPANS:
assert {
@ -329,12 +260,9 @@ async def test_every_emitted_attribute_is_registered(emitted):
@pytest.mark.asyncio
async def test_every_registered_span_is_emitted(emitted):
"""
The direction nothing else catches: the docs must not describe a span that
no longer exists.
"""
# By identity, not by name: a dynamic entry's own string never appears on
# the wire, so comparing strings would be comparing the wrong things.
"The docs should not describe a span that is no longer emitted."
# Compare by identity: the request span's registry name never appears on
# the wire.
resolved = {id(reg.span_for(name, kind)) for name, kind, _ in emitted}
missing = sorted(str(span) for span in reg.SPANS if id(span) not in resolved)
assert not missing, (
@ -346,14 +274,8 @@ async def test_every_registered_span_is_emitted(emitted):
@pytest.mark.asyncio
async def test_every_registered_attribute_is_emitted(emitted):
"""
Every registered attribute, optional or not, must actually be set at least
once by the workload.
`optional` describes whether a reader should expect it on every span, not
whether the code still sets it - so an attribute deleted from the code but
left in the docs has to fail here even when it is marked optional. If a
new attribute only appears in some rare case, extend exercise() to reach
that case.
Every registered attribute, including optional ones, is emitted at least
once. If a new attribute only appears in rare cases, extend exercise().
"""
by_entry = {}
for name, kind, keys in emitted:
@ -381,7 +303,7 @@ def test_registry_has_no_duplicate_names():
def test_registry_entries_are_documented():
"Every entry carries a description - the docs are generated from these."
"Every entry has a description, used to generate the docs."
for span in reg.SPANS:
assert span.description.strip(), f"{span} has no description"
for attribute in span.attributes:
@ -389,7 +311,6 @@ def test_registry_entries_are_documented():
def test_registry_entries_are_usable_as_plain_strings():
"The str subclassing is the whole reason call sites need no wrapper API."
assert isinstance(reg.DB_QUERY, str)
assert isinstance(reg.DB_NAMESPACE, str)
assert reg.DB_QUERY == "db.query"
@ -399,30 +320,19 @@ def test_registry_entries_are_usable_as_plain_strings():
def test_registry_entries_survive_deepcopy_and_pickle():
"""
A copy of an entry is a plain `str`.
These are `str` subclasses whose `__new__` requires the metadata
arguments, so without `__reduce__` `copy` cannot reconstruct one and
raises. That is not academic: the SDK's `ConsoleMetricExporter` renders
data points with `dataclasses.asdict()`, which deepcopies mappings, and
core passes registry entries as metric attribute keys - see
`test_console_metric_exporter_renders_core_metric_points`.
A copied or unpickled entry is a plain str. ConsoleMetricExporter
deepcopies metric attributes, which use registry entries as keys.
"""
for entry in (reg.DB_NAMESPACE, reg.DB_QUERY, reg.M_OPERATION_DURATION):
assert copy.deepcopy({entry: 1}) == {str(entry): 1}
assert type(copy.deepcopy(entry)) is str
assert pickle.loads(pickle.dumps(entry)) == str(entry)
# The metadata still lives on the registered instance itself, which
# is the only place anything reads it.
# The original entry keeps its metadata
assert entry.description.strip()
@pytest.mark.asyncio
async def test_console_metric_exporter_renders_core_metric_points(otel_metrics):
"""
The end-to-end shape of the bug above: a console metrics dump of
Datasette's own points has to survive `dataclasses.asdict()`.
"""
from opentelemetry.sdk.metrics.export import (
ConsoleMetricExporter,
MetricExportResult,
@ -432,8 +342,7 @@ async def test_console_metric_exporter_renders_core_metric_points(otel_metrics):
ds = Datasette(memory=True)
ds.add_memory_database(name)
await ds.invoke_startup()
# One real query, so the dump contains a db.client.operation.duration
# point keyed by the DB_NAMESPACE registry entry.
# Produces a db.client.operation.duration point keyed by DB_NAMESPACE
await ds.get_database(name).execute("select 1")
data = otel_metrics.reader.get_metrics_data()
@ -445,13 +354,8 @@ async def test_console_metric_exporter_renders_core_metric_points(otel_metrics):
def test_every_histogram_declares_bucket_boundaries():
"""
Every histogram must carry explicit boundaries, and only histograms may.
OpenTelemetry's default boundaries start at 5 and are meant for
milliseconds, so a seconds-valued histogram that inherits them records
everything into one bucket. This is a registry self-consistency check, not
a check that the boundaries reached the SDK - for that see
`test_histograms_spread_values_across_buckets` in test_telemetry_metrics.py.
Every histogram declares bucket boundaries, and only histograms do.
OpenTelemetry's defaults are meant for milliseconds, not seconds.
"""
for metric in reg.METRICS:
if metric.kind == reg.HISTOGRAM:
@ -468,13 +372,8 @@ def test_every_histogram_declares_bucket_boundaries():
def test_dynamic_span_lookup():
"""
`dynamic=True` matching, which is how the request span resolves.
The last two assertions are the ones worth having: a dynamic entry must
not swallow a span that does have a registered name, and must not match at
all when the caller supplies no kind - otherwise every unregistered span
in the suite would silently resolve to the request span and the
emitted-but-not-registered direction would stop catching anything.
dynamic=True entries such as the request span match on kind. They never
match without a kind, and never override a registered name.
"""
assert reg.span_for("GET", SpanKind.SERVER) is reg.HTTP_REQUEST
assert reg.span_for("POST /^/(?P<database>[^/]+)$", SpanKind.SERVER) is (
@ -501,27 +400,15 @@ def test_span_and_attribute_lookup():
@pytest_asyncio.fixture
async def emitted_metrics(otel_metrics):
"""
Every (metric name, attribute key) pair produced by a broad workload,
plus the raw set of metric names - the metric-side counterpart of the
`emitted` span fixture above.
Metrics use DELTA temporality (see `otel_meter_provider` in datasette.telemetry_testing), and the
function-scoped `otel_metrics` fixture drains any state left by an
earlier test before yielding, so this collection is not polluted by
other tests in the session - only by other *instances*, which is why the
checks below key everything off attribute names rather than values.
Metric names and (metric name, attribute key) pairs from a broad workload.
Checks use attribute keys rather than values, since other Datasette
instances in the session can also report points.
"""
# The span workload already reaches every synchronous metric except the
# interrupted counter: reads and writes drive db.client.operation.duration
# and datasette.write.queue_wait, and both the suppressed-error probe and
# the custom_time_limit interrupt raise through record_operation_duration,
# setting error.type.
# Reaches every synchronous metric except datasette.sql.queries.interrupted
ds = await exercise()
# datasette.sql.queries.interrupted counts only queries that exceed the
# *configured* limit - a caller opting into a deliberately short budget
# via custom_time_limit (as exercise() does) is excluded by design. So a
# second instance whose configured limit is tiny provides the real thing.
# datasette.sql.queries.interrupted ignores custom_time_limit timeouts, so
# this needs an instance with a low sql_time_limit_ms.
slow_name = _unique("registry_metrics_slow")
slow = Datasette(memory=True, settings={"sql_time_limit_ms": 5})
slow.add_memory_database(slow_name)
@ -533,8 +420,7 @@ async def emitted_metrics(otel_metrics):
"select * from c"
)
# Collect while both instances are still registered, so the observable
# gauges - which observe live instances at collection time - report.
# Collect before closing the instances so the observable gauges report them
otel_metrics.collect()
snapshot = otel_metrics.snapshot
assert snapshot, "no metrics captured - the fixture is not exercising anything"
@ -551,11 +437,8 @@ async def emitted_metrics(otel_metrics):
@pytest.mark.asyncio
async def test_metrics_conform_to_the_registry(emitted_metrics):
"""
Emitted-but-unregistered, via the plugin kit's helper - consumed here
exactly the way a plugin's suite would. Beyond names and attribute keys,
this also asserts each instrument was created as the kind and unit its
registry entry declares, and that `datasette.operation` only ever takes
its declared enum values.
Emitted metric names, kinds, units, attribute keys and enum values match
the registry, using the plugin testing helper.
"""
assert_metrics_conform(
reg.METRICS, emitted_metrics["collector"], scope_name="datasette"
@ -564,7 +447,6 @@ async def test_metrics_conform_to_the_registry(emitted_metrics):
@pytest.mark.asyncio
async def test_every_registered_metric_is_emitted(emitted_metrics):
"Registered-but-never-collected, via the plugin kit's helper."
assert_metrics_covered(
reg.METRICS, emitted_metrics["collector"], scope_name="datasette"
)
@ -572,23 +454,7 @@ async def test_every_registered_metric_is_emitted(emitted_metrics):
@pytest.mark.asyncio
async def test_every_registered_metric_attribute_is_emitted(emitted_metrics):
"""
The direction nothing else catches: the docs must not describe a metric
attribute that no longer exists.
Unlike the span-side attribute check, this does not skip `optional`
attributes. The only optional metric attribute is `error.type` on
`db.client.operation.duration`, and the workload reaches it from two
independent directions: the suppressed-error probe and the
custom_time_limit interrupt in `exercise()`, both of which raise through
`record_operation_duration`. So it is checked like any other attribute
rather than exempted; marking something optional here would opt it out of
verification entirely.
Gauges with no registered attributes (`datasette.sql.threads.limit` and
`.queue_depth`) fall out correctly with no special case: their
`metric.attributes` is empty, so the inner loop makes no assertion.
"""
"Every registered metric attribute, including optional ones, is emitted."
emitted_keys_by_metric = {}
for metric_name, key in emitted_metrics["pairs"]:
emitted_keys_by_metric.setdefault(metric_name, set()).add(key)
@ -596,8 +462,7 @@ async def test_every_registered_metric_attribute_is_emitted(emitted_metrics):
missing = []
for metric in reg.METRICS:
if str(metric) not in emitted_metrics["names"]:
# Not emitted at all - already reported by
# test_every_registered_metric_is_emitted; do not double-report.
# Reported by test_every_registered_metric_is_emitted
continue
emitted_keys = emitted_keys_by_metric.get(str(metric), set())
for attribute in metric.attributes:
@ -610,13 +475,7 @@ async def test_every_registered_metric_attribute_is_emitted(emitted_metrics):
def test_prefix_span_lookup():
"""
`prefix=True` matching, exercised directly.
Core registers no prefix spans - the flag exists for plugin registries
(e.g. a `chat {model}` span family) - so without this the branch in
`span_for()` would be untested code the conformance tests never reach.
"""
"prefix=True matching, which core does not use but plugin registries can."
hook = reg.SpanName("myplugin.hook.", "A hypothetical span family", prefix=True)
spans = reg.SPANS + (hook,)
assert reg.span_for("myplugin.hook.render_cell", spans=spans) is hook
@ -626,7 +485,6 @@ def test_prefix_span_lookup():
def test_exact_match_wins_over_prefix():
"A prefix family can never shadow a span with a registered exact name."
family = reg.SpanName("db.", "Greedy prefix", prefix=True)
spans = (family,) + reg.SPANS
assert reg.span_for("db.query", spans=spans) is reg.DB_QUERY

29
tests/test_telemetry_testing_kit.py
View file

@ -1,8 +1,6 @@
"""
The plugin telemetry kit (`datasette.telemetry_testing` plus the public
registry classes), exercised the way a third-party plugin would use it: a
toy plugin registry, a toy tracer scope, and the kit's own fixtures and
conformance helpers.
Tests for datasette.telemetry_testing and the public registry classes, using
a toy plugin registry and instrumentation scope.
"""
import pytest
@ -54,8 +52,7 @@ def test_conformance_passes_for_a_conforming_workload(otel_spans):
_run_workload()
finished = otel_spans.get_finished_spans()
assert_spans_conform(TOY_SPANS, finished, scope_name=SCOPE)
# Coverage direction needs prefix families seen too - the chat span
# resolves to the CHAT entry despite its variable suffix.
# The chat span matches the CHAT prefix entry:
assert_spans_covered(TOY_SPANS, finished, scope_name=SCOPE)
@ -98,8 +95,7 @@ def test_coverage_catches_a_never_emitted_span(otel_spans):
def test_scope_filter_ignores_other_scopes(otel_spans):
# Core's own spans are in the exporter too; a plugin's conformance run
# must not fail because of them.
# Spans from other scopes, including Datasette's own, are ignored:
other = otel_trace.get_tracer("someone-else", "1.0")
with other.start_as_current_span("not.in.the.toy.registry"):
pass
@ -134,14 +130,9 @@ def test_linked_root_span_kwargs_with_no_current_span(otel_spans):
def test_kit_module_itself_never_imports_the_sdk():
"""
The kit imports the SDK lazily, so a plugin importing it at module
level does not violate the api-only dependency rule.
The kit imports the SDK lazily, so plugins can import it at module level.
conftest.py's pytest_collection_modifyitems() moves this test to the
front of the run by name - if you rename it, rename it there too. Like
every subprocess-spawning test in this suite, running it late crashes
the interpreter on macOS/CPython 3.13 (SIGBUS in fork+exec once the
process holds enough threads) - see the comment there.
conftest.py runs this test first by name. Update it there if you rename it.
"""
assert_package_never_imports_sdk("datasette.telemetry_testing")
@ -159,8 +150,7 @@ from datasette.telemetry_testing import (
toy_meter = otel_metrics_api.get_meter(SCOPE, "0.1")
# Instrument names must be unique per meter for the SDK, so each test mints
# its own via this counter rather than re-registering one name.
# Gives each test a unique instrument name:
_metric_ids = itertools.count()
@ -251,14 +241,13 @@ def test_metrics_covered_skips_optional_attributes(otel_metrics):
error_type = reg.Attribute("toyplugin.error", "Only on failure.", optional=True)
registry = _toy_metric_registry(name, attributes=(OUTCOME, error_type))
counter = toy_meter.create_counter(name, unit="{job}")
counter.add(1, {OUTCOME: "ok"}) # no error attribute - and that is fine
counter.add(1, {OUTCOME: "ok"}) # No error attribute
otel_metrics.collect()
assert_metrics_covered(registry, otel_metrics, scope_name=SCOPE)
def test_metrics_scope_filter_ignores_other_scopes(otel_metrics):
# Core's own metrics are in the reader too; a plugin's conformance run
# must not fail because of them.
# Metrics from other scopes, including Datasette's own, are ignored:
name = f"toyplugin.scoped.{next(_metric_ids)}"
registry = _toy_metric_registry(name)
counter = toy_meter.create_counter(name, unit="{job}")