beads/scripts/ci/_embedded_shard_manifest_lib.py
Julian Knutsen 1137aeb70d
ci(bazel): embedded-Dolt tier 20/5 → 50/15 duration-balanced Bazel shards (F1) (#7156)
* ci(bazel): embedded-Dolt tier 20/5 -> 40/14 duration-balanced shards (F1)

Give the Bazel-only bazel-embedded job its own duration-balanced manifest
blocks for //cmd/bd:bd_embedded_test (20 -> 40 shards) and
//internal/storage/embeddeddolt:embeddeddolt_embedded_test (5 -> 14
shards), decoupled from PR Risk's/main.yml's legacy frozen 20/5-shard
fork/push jobs, mirroring F2's bd_proxied_test precedent.

- Measured every TestEmbedded* / embeddeddolt test's wall time in
  isolation; split 21 cmd-tier and 1 storage-tier single top-level test
  over 180s into disjoint sub-tests (3 irreducible exceptions left:
  TestEmbeddedUpdateConcurrent 475s, TestEmbeddedCloseConcurrent 308s,
  TestLargeBatchApplyStatementCounts712_Embedded 261s).
- New scripts/ci/gen_embedded_{cmd,storage}_shard_manifest.py +
  _embedded_shard_manifest_lib.py generate/check duration-balanced
  manifest blocks (LPT packing, incremental --write/--repack, exact-once
  --check), chosen at N=40/M=14 where the heaviest shard reaches the
  irreducible-test floor.
- Fixed 18 stale legacy 20-shard cmd manifest entries (and 1 storage
  entry) left over from earlier test splits, which would have hard-failed
  every shard of PR Risk's real test-embedded-cmd jobs on their next run.
- New scripts/embedded_shard_count_test.go policy tests tie the
  Bazel-only block's shard count to bazel.yml's check_shard_coverage.py
  argument and catch generator staleness; generalized
  TestBazelRetiredLanesCheckListedTestsRan's live-shard-count dispatch
  from bd_proxied_test-only to a map covering both embedded targets;
  removed TestBazelEmbeddedJobMirrorsEmbeddedTier's now-invalid
  BUILD.bazel-shard-count-equals-PR-Risk-matrix assertion.
- Verified: go test ./scripts/..., bazel test //scripts:scripts_test,
  actionlint, make ci-pr-policy, make bazel-sync-check all pass; full
  local bazel test of both targets at 40/14 shards passes 2/2; mutation
  testing confirms each new/changed policy test fails exactly when it
  should.

See ~/beads-bazel-plan/f1/impl-report.md for full detail.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

* fix(embedded-dolt): skip 712-item statement-count assertion under -race (F1)

TestLargeBatchApplyStatementCounts712_Embedded genuinely costs ~260-300s
under -race (confirmed by solo re-measurement: 298.60s) -- 14014 real SQL
statement round-trips through the race-instrumented in-process Dolt engine,
not an artifact and not splittable (one indivisible ApplyBatchInTx
transaction). It was the storage tier's single-test floor for CI shard
packing.

Apply the exact precedent TestLargeBatchApplyWallClock_Embedded already set
for its 1000-item shape: skip this assertion under raceEnabled (the cost is
race-instrumentation overhead on the Dolt engine's own internal
goroutine/lock machinery, not this test's own logic) and extend nightly.yml's
"Embedded Dolt batch-apply suite (non-race)" step's -run regex so the full
pinned 14014-statement baseline still runs, non-race, every night. Verified
directly: under -race the assertion now skips in 0.00s; without -race it
still runs and asserts Total=14015 (within the existing +/-2 tolerance) in
36.76s.

Updates scripts/ci_workflow_test.go's
TestNightlyFullTestRunsEmbeddedDoltBatchApplySuite to match the extended
regex.

Storage tier's new single-test ceiling under the per-PR -race lane drops
from ~260.7s to 152.38s (TestBatchApplyContract).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

* ci(bazel): correct cmd-tier long-pole durations, re-pack embedded manifolds (F1)

Investigated coordinator item 2 (single-test long poles): TestEmbeddedUpdateConcurrent
(recorded 475.39s) and TestEmbeddedCloseConcurrent (recorded 308.06s), plus
the two other >180s cmd-tier entries from the same capture
(TestEmbeddedInitA 196.94s, TestEmbeddedEpicCloseEligibleReason 192.47s).

Root cause: NOT host contention, NOT backoff/retry sleeps (bdRunWithFlockRetry
contributes negligibly), NOT race-instrumented-subprocess overhead alone.
These numbers were captured via `bazel --nohome_rc test --config=embedded`
with the then-current 20-shard legacy manifest (~10-11 tests/shard) and
`-test.parallel=4`. Go's t.Parallel() starts each parallel test's elapsed
clock when its function is invoked, before the Parallel() call blocks it
waiting for a free one-of-4 slot -- so a test queued behind several other
heavy shard-mates has that queueing wait baked into its own recorded
`--- PASS: TestX (Ys)` time. This is shard-composition-dependent scheduling
delay, not a cost intrinsic to the test.

Confirmed two ways in an isolated clone (not this worktree, per the
no-plain-go-test-from-a-beads-checkout rule):
- Solo (no rivals at all): 151.83s / 106.48s / 63.58s / 66.08s.
- All four together with the real -race -test.parallel=4 flags and exactly
  4 tests requested (so nobody queues -- parallel slots == test count):
  141.34s / 113.80s / 69.79s / 74.63s.
Both agree closely and both are far below the capture-time numbers.

F1's LPT packer already isolates the heaviest tests into shards with few or
no other heavy rivals (TestEmbeddedUpdateConcurrent lands alone in its own
40-shard-block shard), so under the new manifest these tests see little to
no queueing delay in real CI. None of the four needed splitting, parallel
subtests, or raceEnabled-scaled N: once correctly measured, none exceeds the
180s single-test guideline (new corrected weights: 155.0 / 115.0 / 72.0 /
77.0s). Corrected scripts/ci/embedded_cmd_test_durations.json accordingly
(with a doc-comment explaining the methodology and both confirmations) and
re-packed the 40-shard Bazel-only block (--repack --write); --check passes
(exact-once coverage of all 241 discovered tests). The frozen legacy 20-shard
block is untouched and independently re-verified exact-once by simulating
embedded-test-shard.sh across all 20 shards (241/241 tests, no
duplicates/misses/errors).

New cmd-tier single-test ceiling: 179.47s (TestEmbeddedCloseAlreadyClosed,
likely itself somewhat inflated by the same mechanism but already under the
180s line as recorded, so no action needed).

Also updates scripts/ci/embedded_storage_test_durations.json: the prior
commit's raceEnabled-skip for TestLargeBatchApplyStatementCounts712_Embedded
means it now costs ~0s in the per-PR -race lane bazel-embedded actually
packs for (not its real ~260-300s non-race cost, which nightly.yml still
exercises) -- weight set to 1.0 with a doc comment tying it to the skip in
large_batch_apply_measure_test.go. Re-packed the 14-shard Bazel-only block
accordingly; --check passes (exact-once coverage of all 387 discovered
tests, excluding TestMain/TestConformance per that file's existing
convention). The frozen legacy 5-shard block is untouched and independently
re-verified exact-once via embedded-storage-test-shard.sh simulation
(388/388 tests, no duplicates/misses/errors).

Verification: go test ./scripts/... (includes the nightly.yml policy test),
make ci-pr-policy, make bazel-sync-check, actionlint all pass. Real local
`bazel --nohome_rc test --config=embedded //cmd/bd:bd_embedded_test
//internal/storage/embeddeddolt:embeddeddolt_embedded_test`
(--local_test_jobs=8, both targets together, on a shared/noisy sandbox host)
passes both targets; per-shard wall time from each shard's test.xml: cmd/bd
max 218.9s (shard 23/40, min 20.6s), storage max 251.1s (shard 5/14, min
0.98s) -- down from the prior oversubscribed sweep's 783.41s / 351.94s and
well under the single-test-floor-based ~475s projection this slice started
from. Projected tier lane time (max of the two targets, real dedicated CI
runners should do at least this well): ~251s, under the 5-minute target.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

* test(embedded): un-vacuous TestEmbeddedDepB edges, split epic-close test, trim concurrency stress N (F1 review)

Coordinator review follow-ups on the F1 embedded-Dolt resharding slice:

- TestEmbeddedDepB (one of TestEmbeddedDep's two CI-speed split halves) was
  asserting its tree_* subtests against an edgeless graph, passing
  vacuously; restores the same parent-child/blocks edges TestEmbeddedDepA
  builds so those subtests exercise real multi-node traversal (review S2).
- TestEmbeddedEpicCloseEligibleReason split into
  TestEmbeddedEpicCloseEligibleReasonFlag/...Output for CI shard balance,
  matching the project's established single-concern split pattern.
- TestEmbeddedUpdateConcurrent/TestEmbeddedCloseConcurrent's
  issuesPerWorker reduced 5->2: under real concurrent-shard-load
  measurement these were genuine (not queueing-artifact) serialized-write
  costs; 2 preserves every invariant the test checks (duplicate-ID
  detection across all workers, one non-decreasing-list-count check per
  worker) at a fraction of the serialized work.
- large712ShapeName: the 712-shape's race skip and its zero-match guard now
  share one named constant instead of two independent literal shape-name
  strings, so a future rename of the shape can't silently skip the pinned
  statement-count assertion under race (review N2).
- Comment-only fixes (review N5) in files referencing test names since
  renamed/split by this slice's earlier work: close_proxied_integration_test.go,
  graph_apply_test.go, storage_class_proxied_integration_test.go,
  tests/regression/discovery_test.go, plus doc-comment path cleanups (N1:
  drop unverified ~/beads-bazel-plan paths/bead ids in favor of the
  committed duration JSON files and engdocs/TESTING.md).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

* ci(bazel): re-derive embedded-Dolt durations under concurrent load; cap cmd shards at Bazel's 50-shard limit (F1 round 2)

Coordinator review required re-deriving every per-test duration weight from
a single, consistent, CI-realistic measurement (all shards running
concurrently, -test.parallel=4, race-enabled) instead of the prior pass's
mix of crowded-shard captures and solo/4-together spot-corrections, and
required measuring per-shard wall time via BEP testAttemptDuration rather
than a shard's outer test.xml <testsuites time> (which can include a
different, already-finished shard-mate's setup/teardown).

- scripts/ci/embedded_cmd_test_durations.json (242 tests) and
  scripts/ci/embedded_storage_test_durations.json (387 tests) rewritten in
  full from one consistent sweep. Every entry is now a fresh measurement;
  none is a carried-over capture-time value or an unverified estimate. Both
  files' _comment blocks retract the earlier "queueing delay baked into a
  test's own reported elapsed time" theory (testing.go's (*T).Parallel does
  not add wait time to a blocked test's own clock) and document the real
  methodology.
- Only one cmd-tier test remains over 200s (TestEmbeddedListFiltering,
  239.70s), already under the ~240s ceiling; no further top-level splits
  were needed this round.
- New finding: Bazel hard-caps shard_count at 50 per sh_test-like rule
  ("Having more than 50 shards is indicative of poor test organization";
  a native analysis-time limit, not configurable). The previous 88-shard
  plan for //cmd/bd:bd_embedded_test (chosen as the LPT-packed floor before
  this limit was discovered) is reverted to 50, Bazel's real ceiling.
  Repacking the fresh weights at 50 shards gives a heaviest shard of
  409.6s -- above the single-test floor of 239.7s, because 50 bins is not
  enough headroom over the suite's ~20145 measured test-seconds to flatten
  every bin to that floor. This makes the <=5 min lane target unreachable
  for the cmd tier via resharding alone within Bazel's native per-rule
  limit; see ~/beads-bazel-plan/f1/impl-report.md section 3c for the full
  writeup and the two remaining levers (suite-wide speedups or a
  multi-Bazel-rule split), neither attempted in this slice.
- internal/storage/embeddeddolt gets 15 Bazel-only shards (unaffected by
  the 50-shard cap): LPT-packed heaviest shard 225.3s, within target.
- cmd/bd/BUILD.bazel, internal/storage/embeddeddolt/BUILD.bazel,
  .github/workflows/bazel.yml, engdocs/CI_REQUIRED_CHECK_TOPOLOGY.md,
  scripts/embedded_shard_count_test.go, scripts/pr_risk_bazel_coverage_test.go,
  scripts/ci/gen_embedded_cmd_shard_manifest.py updated to the corrected
  50/15 shard counts (and, for the generator, N3: --write without an
  explicit total_shards now errors instead of silently regenerating the
  frozen legacy 20-shard block).
- .github/scripts/embedded-{cmd,storage}-test-shards.txt: re-packed
  Bazel-only blocks at 50/15; frozen legacy 20/5 blocks untouched except for
  N5 fixes to stale entries left by this slice's earlier test splits
  (TestShardScriptsListOnlyRealTests catches this class of bug; both
  legacy blocks reconfirmed exact-once).
- .github/workflows/nightly.yml / scripts/ci_workflow_test.go (N4): the
  embedded-dolt batch-apply nightly step now runs with `if: ${{
  !cancelled() }}` and its own explicit `-timeout 20m`, independent of the
  job-level timeout, with both halves now asserted by
  TestNightlyFullTestRunsEmbeddedDoltBatchApplySuite.

Verified: gofmt -l clean, go vet ./scripts/... clean, go test ./scripts/...
green, both generators' --check pass at 50/15 (and legacy 20/5 remain
exact-once via the frozen-block invariant).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

* test(ci): rename embedded manifest staleness tests out of the ^TestEmbedded skip

The allowlisted go-test leg runs with -skip ^TestEmbedded, so tests named
TestEmbedded*ShardManifestGeneratorNotStale never ran there and the leg
failed with "matched no test that ran".

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Gas City Adopt-PR <adopt-pr@gascity.local>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
2026-10-03 21:30:03 -07:00

168 lines
7 KiB
Python

"""Shared bin-packing/manifest-file machinery for the embedded-Dolt tier's
Bazel-only shard generators (gen_embedded_cmd_shard_manifest.py,
gen_embedded_storage_shard_manifest.py).
This is a near-verbatim extraction of scripts/ci/gen_proxied_shard_manifest.py
(F2)'s LPT packing, incremental --write, and multi-block manifest-file
machinery, generalized so both embedded generators (and any future one) share
one implementation instead of two independently-maintained copies. The
per-suite scripts own discovery (the test-name regex/glob), the duration-file
path and its "seconds_per_init_fallback" semantics, and the manifest header
text; this module owns everything suite-agnostic: pack(), incremental_pack(),
split_blocks()/write_block(), and check_coverage().
See gen_proxied_shard_manifest.py's module docstring for the full rationale
behind incremental --write (merge-conflict avoidance across unrelated PRs)
and --check's exact-once-coverage-only contract (fork-safe: it does not fail
when a fresh repack would reassign shards differently).
"""
import os
import re
def pack(costs, total):
"""LPT bin-pack costs into `total` shards. Returns (shards, loads)."""
order = sorted(costs, key=lambda k: (-costs[k], k))
shards = [[] for _ in range(total)]
loads = [0.0] * total
for name in order:
i = loads.index(min(loads))
shards[i].append(name)
loads[i] += costs[name]
return shards, loads
def split_blocks(text):
"""Split a manifest file's text into its per-total_shards blocks.
A block is a run of leading '#' comment (and blank spacer) lines,
immediately followed by a run of lines that do not start with '#' (shard
assignment lines and the blank lines render() puts between shards) that
ends at the next '#' line or end of file. Returns a list of dicts with
'header' and 'body' line lists and the block's 'total' (parsed from its
first assignment line; None for a block with no assignment lines yet).
"""
lines = text.split('\n')
blocks = []
i, n = 0, len(lines)
while i < n:
header = []
while i < n and lines[i].startswith('#'):
header.append(lines[i])
i += 1
while i < n and lines[i] == '' and not (i + 1 < n and lines[i + 1].startswith('#')):
header.append(lines[i])
i += 1
body = []
while i < n and not lines[i].startswith('#'):
body.append(lines[i])
i += 1
if not header and not body:
continue
total = None
for bl in body:
m = re.match(r'^(\d+) ', bl)
if m:
total = int(m.group(1))
break
blocks.append({'header': header, 'body': body, 'total': total})
return blocks
def read_assignments(path, total):
"""Return {name: shard_number} for path's block matching total_shards ==
total, or None if path or that block does not exist yet."""
if not os.path.exists(path):
return None
for b in split_blocks(open(path).read()):
if b['total'] == total:
assignments = {}
for ln in b['body']:
m = re.match(r'^(\d+) (\d+) (\S+)$', ln)
if m and int(m.group(1)) == total:
assignments[m.group(3)] = int(m.group(2))
return assignments
return None
def incremental_pack(costs, total, existing):
"""Keep every name in `existing` that is still in `costs` on its current
shard; drop names no longer in `costs` (renamed/split/deleted); LPT-place
every name in `costs` not already in `existing` onto whichever shard is
currently lightest. Never moves an already-assigned test. Returns
(shards, loads), like pack()."""
shards = [[] for _ in range(total)]
loads = [0.0] * total
for name, shard_num in existing.items():
if name not in costs or not (1 <= shard_num <= total):
continue # stale, or out of range for a changed total: re-place below
shards[shard_num - 1].append(name)
loads[shard_num - 1] += costs[name]
placed = {n for s in shards for n in s}
new_names = sorted((n for n in costs if n not in placed), key=lambda k: (-costs[k], k))
for name in new_names:
i = loads.index(min(loads))
shards[i].append(name)
loads[i] += costs[name]
return shards, loads
def write_block(path, total, out):
"""Replace path's block for total_shards == total with out (a render()
list), preserving every other block and its position. Appends out as a
new trailing block if path has none for this total yet."""
text = open(path).read() if os.path.exists(path) else ''
blocks = split_blocks(text) if text else []
result = []
replaced = False
for b in blocks:
if b['total'] == total:
result.extend(out)
replaced = True
else:
result.extend(b['header'])
result.extend(b['body'])
if not replaced:
if result and result[-1] != '':
result.append('')
result.extend(out)
with open(path, 'w') as fh:
fh.write('\n'.join(result).rstrip('\n') + '\n')
def check_coverage(path, total, universe, test_name_desc, script_name):
"""Return None if path's committed block for total_shards == total lists
every name in `universe` (the currently-discovered test set) exactly
once, with no stale or duplicate names; otherwise an actionable error
string. Does NOT check the block's shard *assignments* — see this
module's docstring for why (incremental --write)."""
if not os.path.exists(path):
return f'{path} does not exist'
for b in split_blocks(open(path).read()):
if b['total'] == total:
names = []
for ln in b['body']:
m = re.match(r'^(\d+) (\d+) (\S+)$', ln)
if m and int(m.group(1)) == total:
names.append(m.group(3))
universe_set = set(universe)
counts = {}
for n in names:
counts[n] = counts.get(n, 0) + 1
dupes = sorted(n for n, c in counts.items() if c > 1)
stale = sorted(n for n in counts if n not in universe_set)
missing = sorted(n for n in universe_set if n not in counts)
if not dupes and not stale and not missing:
return None
weights_guess = 'duration' if any('duration' in h for h in b['header']) else 'inits'
msg = [f"{path}'s {total}-shard block does not have exact-once "
f'coverage of the currently-discovered {test_name_desc} test set.']
if missing:
msg.append(f'missing ({len(missing)}): {", ".join(missing)}')
if stale:
msg.append(f'stale/unknown ({len(stale)}): {", ".join(stale)}')
if dupes:
msg.append(f'duplicated ({len(dupes)}): {", ".join(dupes)}')
msg.append(f'Run: python3 {script_name} {total} --weights={weights_guess} --write')
return '\n'.join(msg)
return f'{path} has no block for total_shards={total}'