pomazanbohdan/memory-mcp-1file

🔴 Critical: Dockerfile Was Running Ancient Binary

Dockerfile.release hardcoded VERSION=0.1.4 — all Docker containers were running the ancient v0.1.4 binary. None of the v0.2.x fixes (crash fix, dimension migration, idle timeout, etc.) were reaching Docker deployments.

Memory Optimizations (OOM Kill Prevention)

Container was killed by OOM Killer (exit code 137) ~15 seconds into indexing. Root cause: producer-consumer imbalance — indexer pushes chunks at disk I/O speed (~1000/sec), but BERT inference processes ~8/sec. Queue fills instantly, tensors accumulate.

Changes

| Component | Before | After | Memory Impact |
|-----------|--------|-------|---------------|
| Embedding queue | 1000 | 64 | Proper backpressure |
| Batch size | 8 | 2 | Tensor memory -4x (~50→12MB) |
| Moka RAM cache | 10K entries | 2K entries | -24MB |
| Queue watermark | 80% | 50% | Earlier throttling |
| Throttle delay | 50ms | 500ms | Stronger backpressure |
| Tensor lifecycle | Implicit | Explicit drop() | No accumulation |

Estimated peak RSS: ~700-900MB → ~400-500MB

Files Changed (9 files, +23/-17 lines)

• Dockerfile.release — VERSION 0.1.4 → 0.2.12
• src/embedding/adaptive_queue.rs — capacity, watermark, throttle
• src/embedding/worker.rs — batch size 8 → 2
• src/embedding/store.rs — moka cache 10K → 2K
• src/embedding/engine.rs — explicit tensor drops
• src/main.rs — mpsc channel 1000 → 64
• Cargo.toml / npm/package.json / Cargo.lock — version bump

Tests

• 69 passed, 0 failed ✅