You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
find_queries_by_text can still stop at its budget with no span finished, so the search returns no rows #9219
Build 3.13.0-nightly.20261121.714. #9150 (closed by #9154) asked for two things: a search cut by the budget returns what it found, and it names the span it searched. The second is done: every reply carries searched with complete, searched_from and searched_to. The first does not hold when the budget runs out before the first span finishes. The reply then says nothing was searched, has complete: false with searched_from and searched_to null, and returns no rows, which is the answer #9150 set out to remove.
Seen on both SQL Server stores tried:
All servers, a common pattern (no server_name): stopped at the 4 s budget with nothing searched, every time it was tried.
One server, a rare pattern, no database: with limit 3 it stopped with nothing searched twice in a row; with limit 2 it answered ok and complete: true in under a second with rows; with the default limit it stopped after the newest few hours of the raw tier, with rows and the covered span named. So whether the first span finishes in time changes from call to call with the same pattern, server and window.
A user who gets an empty stopped reply cannot tell a pattern that is not there from one the search never reached, and the only advice it gives is to name a database, which the all-servers shape cannot take.
Proposed
Size the first span to the budget: start with the newest slice of the raw tier and widen it only while time remains, so a stopped search always returns the newest matches with the span it covered. The all-servers and one-server shapes share this.
For the all-servers shape, when the digest probe has matched texts, read the samples newest first and stop at the budget, rather than joining the whole raw window.
Keep searched as it is.
Tests
All-servers search for a common pattern on a seeded store with a long history and a budget the full window cannot meet: returns at least the newest matches, searched_from is set, complete is false.
One-server search for a rare pattern with no database at limits 1, 2, 3 and the default, under the same tight budget: none stops with searched_from null.
A search that finishes reports complete: true and the full window, as today.
A pattern that is not there answers empty and complete: true, not stopped.
#9230 is merged to dev. This issue stays open until the all-servers search is timed again on a large store, after the nightly build that carries the change. It passes when an all-servers search for a common pattern returns rows with searched_from set, instead of stopping at the time budget with nothing.
Activity
Test the newest spans' text matches against the spans' own digests in…
Allow the lookup-span test's real-clock seeds in the LiveClock source…
Merge pull request #9230 from erikdarlingdata/fix/9219-text-search-fi…
erikdarlingdata commentedon Nov 21, 2026
Claude posting for Erik Darling.
Fixed on dev by #9230.
erikdarlingdata commentedon Nov 21, 2026
Claude posting for Erik Darling.
#9230 is merged to dev. This issue stays open until the all-servers search is timed again on a large store, after the nightly build that carries the change. It passes when an all-servers search for a common pattern returns rows with
searched_fromset, instead of stopping at the time budget with nothing.