Storage
Eviction savings, dev-store throughput, repair, and SQL bounds over upstream denoland/celld.
Patches to what Cello stores and how it reads it back: cheaper evictions, a single-writer dev store, compaction and crash repair, and bounds on SQL and page cache. Request-count savings live on the Efficiency page; each entry names its fork patch number.
Bucket-posture writes and eviction
0054— A bucket-posture eviction skips visibility LISTs when this residency itself uploaded into the prefix: a successor restores by listing, so the uploader’s own success response is proof enough. A resumed local copy, recovered tails, bundles, and the L0-chain fallback still list.0056— Concurrent bucket-posture writes on one cell share ownership reads: a gated write joins a read in flight on its cell and epoch when the issuing gate’s cover reaches its position, and issues its own read at once otherwise. No write waits longer than before; fleet proofs never share reads.0057— An eviction deletes the per-cell L0 and L1 objects its handoff snapshot supersedes: after the snapshot PUT succeeds, a background task deletes this epoch’s journaled upload keys (max txid at or below the snapshot’s) with one single-keyDeleteObjectper key — never bulkPOST ?delete— at bounded concurrency. The eviction returns at once; failed DELETEs leave restore-invisible garbage for epoch GC. Journal capped at 8192 entries per cell epoch.0061— Follow-up: the eviction guard proves a journaled epoch with one HEAD on the newest journaled key instead of a prefix LIST, falling back to the LIST on a miss, an error, or an empty journal.
Local store and crash behavior
0098— Compaction retains live-named epoch files: after an L1 fold the node keeps the newest covered L0 file thatpos()and verify still read back, and each activation plus each startup drops live epoch dirs a newer local epoch supersedes (upstream #311).0099— The dev store funnels every PUT through one writer thread, one connection, and one queued batch per transaction. Concurrent PUTs no longer lose the SQLite lock lottery intobusy_timeoutsleeps: an idle batch costs one commit like today’s single PUT, a loaded batch shares one fsync, and each row keeps its own savepoint so create/update conflicts still fail alone (upstream #310). Local dev-store only; fleet and bucket durability paths untouched.0100— The 0.5.1 facet-image import is repaired: the check is no-user-data instead ofpage_count == 0, which never fired under replication, and the decision is stamped inuser_versionso a later wipe never resurrects it (upstream #292).0105— Unclean restarts prune live-named epochs (see Fleet).0122— Local LISTs read from a covering index over a prefix range instead of scanning the wholeobjects.sqlite3: the open createsobjects_listingand bounds the scan to the prefix, with the Rust-side prefix check as backstop. No new knob; dev-store only (upstream #265).
Bounds and budgets
0066— A per-connection SQLite page-cache budget:finish_opensetsPRAGMA cache_size=-<kb>fromCELLO_SQLITE_CACHE_KB(default 2000, the stock ceiling: no behavior change at merge), validated at startup.0102— Hydration counting is O(1): the hydration set carries an exact cut-relative count maintained on every insert and truncation, so a faulted read no longer walks the whole cut (upstream #270).0128— Per-statement SQL time is bounded byCELLO_SQL_STATEMENT_MS: when set, each statement is interrupted past the bound and fails with a catchable timeout error that rolls the statement back; unset means unbounded, the historical behavior. Streaming cursors share their call’s single deadline, and a timeout flag from an armed run never rewrites a later unarmed error (upstream #274).