Status (at mirror time): Open (tracking) • Severity: N/A (refactor)
One-liner: Drop Redis in favor of SQLite + in-process state. Deferred refactor
Cross-ref: this fits the Phase 0 setup-simplicity theme below
(atlas-bootstrap, prebuilt images, doctor). If Phase 0 lands first
Redis goes from "annoying setup step" to "annoying daemon to remove
later"; doing PC-015 before Phase 0 means Phase 0 has one fewer thing
to install.
- Status: Open (tracking only — no implementation in this sprint)
- Severity: N/A (refactor for setup simplicity, not a bug)
- Location:
geometric-lens/cache/, geometric-lens/router/,
geometric-lens/main.py (task queue + metrics)
Redis is currently load-bearing for the Pattern Cache (sorted sets,
JSON blobs, hit/miss counters), the co-occurrence graph, the router's
Thompson sampling state, the task queue, and metrics. Audit shows the
actual Redis surface is purely key-value + sorted sets + lists — no
pub/sub, no streams, no Lua, no clustering. Every command in use maps
cleanly to SQLite (FTS5 even gives BM25 natively if we ever want to
retire the in-memory PatternMatcher).
Costs Redis adds today:
- An extra service in
docker-compose.yml with healthcheck and
network dep.
- A host kernel prereq (PC-011 /
vm.overcommit_memory=1).
- AOF persistence config and tuning surface.
- A synchronous network hop on every cache op.
- Version skew between client lib and server.
What porting would touch:
cache/pattern_store.py → SQLite tables: patterns(id, json, tier, score, last_modified) with indexes on (tier, score DESC).cache/co_occurrence.py → cooccur_edges(from_id, to_id, count).router/feedback_recorder.py + router/route_selector.py → either
a router_state table or a JSON file with file-lock (state is
small).main.py task queue + metrics → either drop (consumer audit needed
per PC-014) or move to SQLite + JSONL.
Estimated 4 files, ~200-400 LOC. Not in scope for the current sprint;
keep Redis intact. Revisit once the cache fixes (PC-001..PC-010) have
proven out and the topology question (PC-014) is resolved.
Fix
Deliberately deferred. No implementation planned for this sprint.
Mirrored from local ISSUES.md. ISSUES.md is gitignored; this issue is the canonical public record. PC-### numbering preserved for cross-references.
Status (at mirror time): Open (tracking) • Severity: N/A (refactor)
One-liner: Drop Redis in favor of SQLite + in-process state. Deferred refactor
geometric-lens/cache/,geometric-lens/router/,geometric-lens/main.py(task queue + metrics)Redis is currently load-bearing for the Pattern Cache (sorted sets,
JSON blobs, hit/miss counters), the co-occurrence graph, the router's
Thompson sampling state, the task queue, and metrics. Audit shows the
actual Redis surface is purely key-value + sorted sets + lists — no
pub/sub, no streams, no Lua, no clustering. Every command in use maps
cleanly to SQLite (FTS5 even gives BM25 natively if we ever want to
retire the in-memory PatternMatcher).
Costs Redis adds today:
docker-compose.ymlwith healthcheck andnetwork dep.
vm.overcommit_memory=1).What porting would touch:
cache/pattern_store.py→ SQLite tables:patterns(id, json, tier, score, last_modified)with indexes on(tier, score DESC).cache/co_occurrence.py→cooccur_edges(from_id, to_id, count).router/feedback_recorder.py+router/route_selector.py→ eithera
router_statetable or a JSON file with file-lock (state issmall).
main.pytask queue + metrics → either drop (consumer audit neededper PC-014) or move to SQLite + JSONL.
Estimated 4 files, ~200-400 LOC. Not in scope for the current sprint;
keep Redis intact. Revisit once the cache fixes (PC-001..PC-010) have
proven out and the topology question (PC-014) is resolved.
Fix
Deliberately deferred. No implementation planned for this sprint.
Mirrored from local
ISSUES.md. ISSUES.md is gitignored; this issue is the canonical public record. PC-### numbering preserved for cross-references.