| Comment: | Merge pull request #501 from danmestas/perf/submanifest-node-memo
perf(manifest): memoize parsed sub-manifests by hash |
|---|---|
| Timelines: | family | ancestors | descendants | both | main | trunk |
| Files: | files | file ages | folders |
| SHA3-256: |
d82fddda6aee1206dd5c6d7f379035e8 |
| User & Date: | [email protected] on 2026-10-26 00:11:36.000 |
| Other Links: | branch diff | manifest | tags |
|
2026-10-26
| ||
| 00:59 |
test(conformance): hand each case its subtest's t; cancel stale Race runs
check runs each case as a subtest, but the suites passed the case the parent's t, captured in a closure. A fixture that skips or fails, as the fossil-built R-card fixture skips when fossil(1) is absent, then called SkipNow or FailNow on the parent from inside the subtest, which Go fails as "subtest may have called FailNow on a parent test". The Race workflow installs no fossil and failed this way in TestSparseCone. check now hands conforms the subtest's t, and all 150 call sites take it, so a case's skip or failure ends that case alone. TestCheckSkipEndsOnlyTheCase guards the harness. A Race run takes most of an hour and merges land faster, so a new run on a ref now cancels the older one. check-in: 43df2fd2c2 user: [email protected] tags: fix/conformance-subtest-t, main, trunk | |
| 00:54 |
Merge pull request #502 from danmestas/docs/sdk-reference-check
docs: regenerate the SDK reference and check it in CI check-in: 51d3b9cf5f user: [email protected] tags: main, trunk | |
| 00:46 |
docs: regenerate the SDK reference and check it in CI
The gomarkdoc SDK pages had drifted about a thousand lines from the code: make docs-gen-sdk was never part of docs-gen, so nothing regenerated them and the docs check did not compare them. ReadBlob, SetContentCache, ListFilesIn and OriginNamer were missing from the libfossil page. scripts/gen-sdk-docs.sh now renders them, running gomarkdoc v1.1.0 through go run so CI needs no install, then adds the frontmatter. It honours OUT_DIR, gen-docs.sh calls it, and check-docs-generated.sh compares the result, so a stale SDK page fails CI like any other generated input. make docs-gen-sdk calls the same script. llms.txt and llms-full.txt are regenerated from the refreshed pages. check-in: c12b02189f user: [email protected] tags: main, trunk | |
| 00:11 |
Merge pull request #501 from danmestas/perf/submanifest-node-memo
perf(manifest): memoize parsed sub-manifests by hash check-in: d82fddda6a user: [email protected] tags: main, trunk | |
| 00:02 |
perf(manifest): memoize parsed sub-manifests by hash
draft-fossil-submanifests-00 §7.4: enumeration results SHOULD be cached keyed by artifact name, so a rebuild parses each distinct sub-manifest once. The tree walk's loader now serves a resolved node from, and keeps it in, a new byte-bounded LRU memo on content.Cache (DerivedSized), keyed by the node's hash and charged its bytes, within a quarter of the cache's budget. The existing 32-entry Derived FIFO is sized for a few parsed check-ins; a tree's working set is every distinct node, which would thrash it and evict those check-ins. Only a node that resolved is kept, so a phantom is looked up afresh and fails the walk as before. Depth, cycle and node-count bounds stay per walk. The memo is dropped with Derived on Invalidate and Clear, and the Repo's cache is replaced at every write. Listing a tree that differs from one already read in one leaf now parses its manifest and that leaf; listing every check-in of zit's recorded tree history parses 48 sub-manifests instead of 109. SM-7.4-1 is covered. Tracked in #493. check-in: 94f61db228 user: [email protected] tags: main, trunk | |
| 00:01 |
Merge pull request #500 from danmestas/feat/list-files-in-cone
feat: list a check-in's files in a cone check-in: e90c972c40 user: [email protected] tags: main, trunk | |