Repository navigation
sessions: relation refresh rereads the whole session per appended message #2607
Description
Summary
With the git-convergence and derived-anchor full scans from #2521 removed, the largest remaining reads under a steady ingest stream come from session-temporal relation projection. It re-reads every occurrence of the refreshed session generation together with its observations.observation_json and retrieval_anchors.anchor_json (the SELECT occurrence.occurrence_id, occurrence.retrieval_anchor_id, anchor.anchor_json, occurrence.message_id, occurrence.agent_id, occurrence.projection_output_ordinal, occurrence.knowledge_at, occurrence.valid_time_json, observation.observation_json ... page in crates/tracedecay-session-temporal-store). Each appended message therefore costs O(size of its session), not O(new bytes).
Evidence (isolated profile, synthetic corpus: 31k project messages, 8 live transcripts receiving one message every 0.5 s)
- 183 runs in a ~3 minute window read 507 MB: 2.8 MB average and 9.7 MB maximum per run, the top statement by bytes.
- The largest live transcripts in the stream hold about 100–400 messages. Operator sessions reach thousands, so the per-message cost grows with session length.
Expected
A refresh that admits k new messages reads O(k) occurrence and observation rows plus bounded generation metadata. Carry forward the settled prefix instead of rehydrating it. Found while root-causing #2521.
Activity
ScriptedAlchemy commentedon Sep 29, 2026
Status after #2639 (fd451f3) and #2646 (9606347): not fixed. This is now the largest store-scaled read on the stream path.
Root cause (source). Every refresh completion works on the whole session generation:
GENERATION_COPY_STATEMENTS: occurrences, turns, derived evidence, FTS).observation_jsonandanchor_json(rebuild_candidate_session_relations→reconstruct_session_relation_projection→reconstruct_occurrences).validate_final_projection_receipt,copy_identity_coverageviareconstruct_logical_copy_relations).So each refresh is O(session) in reads and writes by construction. "Carry the settled prefix forward" needs a change to the generation model, not a query fix:
session_occurrencescolumns, notobservation_json.Evidence (isolated profile, 1,200-transcript synthetic corpus, 8 live transcripts at one message per 0.5 s, per-statement SQLite profile). Two consecutive 300 s windows on the same store; each window adds about 75 messages to each live session:
The new work per message is the same in both windows. Per-message cost doubled because the sessions got longer. The relation reconstruct query alone went from 662 MB (87 runs) to 1.23 GB (108 runs). Hotpath
session_temporal.txn.complete_refreshwent from avg 434 ms to 674 ms, andsession_temporal.rebuild.relationsfrom 102 ms to 151 ms.Leaving this open for that design; #2521 stays open on it.
perf(sessions)!: extend session generations, don't copy them (Scripte…