Timeline

afterDephantomize crosslinked every artifact it discovered through crosslinkSingle, which opened its own write transaction per artifact and passed an explicit nil content cache into crosslinkCheckin. Both choices suit an isolated artifact and neither suits the cascade: one phantom fill unblocks hundreds or thousands of related artifacts that share delta bases and parent manifests, so the walk paid a transaction commit per artifact and re-expanded content a sibling had just expanded. A CPU profile of round 7 of a fossil-corpus clone showed exactly that pair -- _vdbeCommit plus expandChain/decompress reached through crosslinkSingle -- as the remaining 300+ seconds after the statement-cache fix.

crosslinkSingle had no callers outside the cascade, so it is replaced rather than kept alongside a second strategy. cascadeLinker buffers the artifacts the walk discovers and links them through linkArtifact -- the same type switch the whole-repository sweep uses, so the two paths no longer carry near-duplicate switches -- on one content.Cache for the whole cascade, committing once per crosslinkBatchSize artifacts instead of once per artifact. The sweep's own batching, cache, and tuning constants are untouched.

The walk's delta-child query filters on committed mlink rows, so deferring writes past it could change which artifacts it visits. Orphans are the only discovery that query can observe -- delta.rid is a primary key, so every other buffered artifact was itself produced by that query and cannot come back -- so the buffer flushes after an orphan batch and before the query, preserving the original order: orphans crosslinked, orphan rows cleared, children queried.

Failure handling keeps what per-artifact transactions bought. An artifact that cannot be expanded, parsed, or linked is logged with the same wording and skipped while the rest of the batch commits, and each artifact's writes are wrapped in a savepoint so a link that fails partway leaves no partial event/plink/mlink/tagxref behind. Cancellation is unchanged in the walk and newly observed inside a flush on the sweep's stride; an open batch is abandoned rather than committed, so committed batches stand and the next Crosslink sweep re-selects whatever still has no event row.

Tests: a characterization test pins the cascade's derived rows and visit count over a graph covering both sides of the mlink filter -- an orphan that writes mlink rows and one that does not, and so is still visited twice -- against the dump the per-artifact shape produced. A regression guard asserts one commit and one expansion per artifact with the parent manifest served from cache for a 300-artifact burst, plus tests for whole-batch flushing, savepoint isolation, and mid-flush cancellation. check-in: bfb2ad3aa4 user: [email protected] tags: trunk

The hot per-artifact queries in the phantom-fill crosslink cascade — the content/blob-by-rid lookup and the per-node orphan/delta-child lookups in afterDephantomize — re-tokenized, re-parsed, and re-compiled their SQL text on every call because db.Querier delegated straight to the raw driver each time with no statement reuse. On a large repository (67k+ artifacts) the SQLite parser/compiler dominated CPU during the slow clone stretch.

Add a statement cache to both Querier implementations, keyed by exact query text, with lifetimes matched to each backing type:

- *DB caches *sql.Stmt prepared from its long-lived *sql.DB pool (safe for concurrent use, manages its own connections). Statements are prepared lazily on first use, reused thereafter, and closed in DB.Close. - *Tx caches *sql.Stmt scoped to a single *sql.Tx; the cache is torn down (all statements closed) when the transaction commits or rolls back, so no statement outlives its transaction or leaks across a clone/sync session.

Exec/QueryRow/Query transparently get-or-prepare the cached statement — no signature or caller-visible behavior change. QueryRow preserves database/sql's deferred-error contract by falling back to the raw path on prepare failure so the error still surfaces from Scan. Multi-statement strings (DDL setup batches) bypass the cache and run through the raw driver path, since a prepared statement compiles only one statement.

Benchmark (Apple M4 Pro, modernc driver): the blob-by-rid lookup drops from 4779 ns/op to 2289 ns/op (~2.1x). Cache reuse, transaction-boundary cleanup, and multi-statement bypass are covered by new tests in db/stmtcache_test.go. check-in: 218ebaec74 user: [email protected] tags: trunk