mirror of
https://github.com/windmill-labs/windmill.git
synced 2026-08-19 00:02:03 +00:00
Ruben Fiszel
577ceeee86
* [ee] perf(audit): re-anchor S3 audit export on enable + opt-in backfill
The S3/GCS audit-log export's steady-state query filters by `age(xmin)`
(unindexable), so the only scan bound is the timestamp floor. On a fresh
enable the floor was epoch, and on a re-enable the cursor resumed from its
pre-disable position — either way the first run scanned the whole
`audit_partitioned` table. Under a `statement_timeout` (e.g. Aiven) that scan
never completes: the cursor never advances, nothing is exported, and the
repeated full scans saturate the database.
Re-anchor on enable (EE companion, windmill-ee-private#634):
- New trigger migration records a recent timestamp floor instead of the epoch
sentinel and `DO UPDATE`s the cursor to the current snapshot xmin on
re-enable, so the export always resumes from ~now and never rescans history.
Includes a one-time fixup for legacy epoch-sentinel checkpoints on upgrade.
Opt-in historical backfill (new `audit_logs_s3_backfill` module + endpoints):
- Exports a chosen `[from, to)` window on demand, scanning strictly by
`timestamp` (the partition key) in bounded keyset pages — each query is an
index scan capped at one page (verified via EXPLAIN: later partitions
`never executed`, ~11ms/page), so it stays well under any statement timeout
regardless of window size. Writes alongside the steady-state objects under
logs/audit/, without touching the xmin cursor.
- POST /settings/audit_logs_s3_backfill {from,to} (super-admin + Enterprise),
GET /settings/audit_logs_s3_backfill_status.
Also repurposes the status endpoint's `bootstrapping` flag to mean "draining a
backlog" (the cursor is capped and catching up), and updates the setting
description to point operators at the backfill.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(audit): heartbeat backfill lease per object; bump EE ref
Address review (cubic): persist progress (refreshing the lease heartbeat) after
every object PUT in the backfill page loop, not only once per page, so the gap
between heartbeats stays well under STALE_HEARTBEAT_SECS even on slow uploads
and another replica can't re-claim mid-page and run a concurrent backfill.
Bumps ee-repo-ref.txt to pull in the EE test-race fix (folding the backlog-drain
regression into the single audit e2e test).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(audit): reject unstable backfill windows; bump EE ref
Address review (P1): the backfill keyset-pages over rows visible at scan time
and declares completion when the scan runs dry, but a row's `timestamp` is its
inserting transaction's `xact_start`. A window whose upper bound is recent or in
the future could silently omit a transaction that started inside `[from, to)`
but commits after the scan passed that timestamp. `try_start` now rejects any
`to` newer than the oldest in-flight `xact_start` (everything strictly older
than the oldest running transaction is committed and stable), using the same
trustworthy stats gating as the exporter's floor (restricted role / 2PC → a
7-day-old cutoff).
Bumps ee-repo-ref.txt for the EE monotonic-checkpoint fix.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(audit): re-anchor legacy epoch checkpoints instead of synthetic floor
Address review (P1): the legacy-checkpoint fixup stamped last_oldest_inflight_ts
to now()-7d while leaving the old last_xmin in place. On an instance that
enabled export on the old code >7 days ago and got stuck before the first
successful batch, the next run would filter post-enable rows older than 7 days
out via `timestamp >= ts_floor` while still advancing last_xmin over the
interval — silently dropping them (the same floor-vs-cursor loss class fixed
elsewhere in this PR), and contradicting the "nothing committed after enabling
is skipped" guarantee.
A stuck epoch-sentinel checkpoint cannot be safely resumed (its backlog can be
arbitrarily old, so any recent floor prunes rows the cursor then skips, and an
epoch floor reintroduces the full scan). Re-anchor it to the migration's current
snapshot xmin instead — exactly like a fresh enable — so the export resumes
cleanly from ~now and the never-exported pre-upgrade window is recovered via the
opt-in backfill rather than silently dropped. Reword the setting description so
it no longer implies the disabled/legacy window is covered by the cursor.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(audit): end-to-end integration tests for the object-store backfill
The backfill previously had only SQL-level/EXPLAIN validation. Add real
integration tests (in-memory object store, sqlx::test) exercising the public
path:
- backfill_exports_window_in_pages: with the page size forced to 2 rows, a
settled 3-day window is exported across multiple keyset pages; asserts every
in-window row lands exactly once, rows outside [from,to) are excluded, a day
that straddles a page boundary yields more than one object, progress counts
match, and a re-run is idempotent (deterministic keys overwritten, no dupes).
- backfill_rejects_unstable_window: a future/live `to` is rejected as unstable,
a window safely in the past is accepted.
Adds a test-only PAGE_ROWS override so multi-page behaviour is exercised with a
handful of rows.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* docs(audit): note backfill scope is audit_partitioned only
Make explicit that, like the steady-state export, the backfill reads only
audit_partitioned; the pre-partitioning `audit` table is intentionally out of
scope (not a missed case).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(audit): reject backfill windows before the partitioned boundary
Address review (Codex P1): the backfill reads only audit_partitioned, but
pre-partitioning history lives in the legacy `audit` table (still read by audit
list/get via UNION ALL, and retained for the configured period — 365 days by
default on EE). Since the setting text points operators at this API for
"pre-existing history", a window overlapping legacy rows would report completion
while silently omitting them.
Per the decision to not export the legacy table, reject instead of silently
omit: try_start now rejects a `from` earlier than the oldest audit_partitioned
timestamp (every legacy row predates the partition cutover, so a `from` at/after
that boundary can never overlap them). Reworded the setting text to scope the
backfill to the partitioned era. Added a regression test, plus an RAII guard
(cubic P2) so the test-only globals are restored even if an assertion panics.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(audit): backfill object keys per-window; require trustworthy settled cutoff
Address review (two P1s):
- Object-key overwrite loss: keys were `dt=<day>/audit_backfill_<min_id>.ndjson`.
A narrower, overlapping backfill can start a day's page at the same first row
(same min_id) but hold fewer rows, and `put` would overwrite a broader run's
object — silently dropping the rows only that object held. Include the
requested window in the key so different ranges write disjoint objects (same
window re-runs stay idempotent; consumers dedupe overlapping rows by id). New
regression test (verified red→green).
- Untrustworthy settled cutoff: when min(xact_start) isn't trustworthy (role
lacks pg_read_all_stats/superuser, or a prepared 2PC txn exists), the old
now()-7d fallback could still let an old transaction commit rows inside an
accepted window after the scan, so a "complete" backfill silently missed them.
Since a backfill asserts completeness, reject in those cases instead of
falling back. (The continuous exporter keeps its 7-day fallback — it only
claims bounded lag.)
Also makes the tests robust under the parallel runner: run_backfill takes the
store as a param, so tests pass a local in-memory store (no global
OBJECT_STORE_SETTINGS race) and serialize on the PAGE_ROWS override.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(audit): reject backfill overlapping legacy table; regen deref openapi; trim migration comment
Address review (1 P1 + 2 P2):
- Empty-partition backfill (P1): the min(audit_partitioned) guard no-ops when
audit_partitioned is empty, so an upgraded instance with legacy `audit` rows
but no partitioned rows yet would accept a window and complete with zero rows,
silently omitting the legacy rows. Check the legacy `audit` table directly:
reject any window that overlaps a legacy row (subsumes the boundary check and
covers the empty-partitioned case). Test updated accordingly.
- openapi-deref (P2): regenerate openapi-deref.yaml/json (served via include_str!)
so /openapi.{yaml,json} expose the new backfill endpoints.
- Migration comment (P2): trim the PR-history narration to the durable
constraints (why a recent floor and a monotonic cursor are required), per
AGENTS.md.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* chore: update ee-repo-ref to b821fecccbcba2efed544890576bf2b84321d70d
This commit updates the EE repository reference after PR #634 was merged in windmill-ee-private.
Previous ee-repo-ref: 6b191b77aabcf77658ad4f9031576e0d7b66bf89
New ee-repo-ref: b821fecccbcba2efed544890576bf2b84321d70d
Automated by sync-ee-ref workflow.
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: windmill-internal-app[bot] <windmill-internal-app[bot]@users.noreply.github.com>