MESA Core · v0.9.3
Performance & limitations
This guide has equivalent English and Turkish versions. The source sections behind this page are at the bottom.
Purpose and boundaries#
No general latency, throughput, accuracy or hallucination-elimination guarantee is established. Local deterministic fixtures verify contracts; mock vectors do not measure semantic quality. Graph limits bound work, not completeness. Context budgets apply to formatted evidence; conservative offline counting can underfill. Read dated benchmark conditions rather than treating historical numbers as current provider results.
Continue with the maintained contracts#
Use the pinned source reference below for complete tables and operator details. Related guides: architecture, API, operations and release limits.
v0.9.3 measured readiness#
The tagged changelog reports full readiness for a 4,999,000-byte document in 245.20 s, versus 718.27 s before (2.93× faster), with models off and graph on: all 4,882 mutations COMMITTED, SQL and GRAPH COMPLETED, VECTOR SKIPPED, lossless chunk readback. The separate final Opus Round 20 check reports 239.35 s (3.00×) on dbf113b. These local measurements are not a service guarantee.
v0.9.3 known issues#
- Per-minute and daily rate limits are not enforced for
/v4/*and/control/*. - ASCII
yargitaydoes not matchYargıtay; index-side Turkish folding remains deferred. - Kùzu 0.11.3 can write uninitialized memory into graph page tails. Raw graph files may contain fragments of text handled by the process even though memory text is not stored in logical graph properties (checked by query). Treat the graph directory as sensitive like SQLite.
See the tagged changelog. Production NO-GO remains. Historical 0.9.2 search measurements are retained in release notes; no 0.9.3 retrieval benchmark was run.
Source reference for this page
The sections of the upstream file that back this page, unchanged. Source section links open the appendix automatically.
Performance and limitations#
No v0.8.0 throughput, recall, accuracy, latency or 10–20 GB corpus-capacity result
has been measured in this documentation release. A unit/fixture smoke is not a
production soak, and bounded graph traversal does not guarantee complete recall.
Historical benchmark and audit documents have been retired from the working
tree. Earlier reports in Git history describe their own fixtures and revisions,
not current capacity or a completed 24-hour release gate. Use
release requirements to evaluate evidence for a specific commit;
missing raw results or run artifacts are unavailable evidence.
VectorEngine maintenance defaults to 128 mutations and cleans old versions
older than seven days. Compaction, deduplication and executor cleanup mitigate
growth; they do not establish bounded long-soak RSS for every workload. Watch
fragments, version metadata, RSS, disk, mutation latency and maintenance failure
counts. Retained projection generations/backups are not automatically deleted.
Queue admission defaults are 50,000 pending records / 2 GiB, with tenant-aware
bounds and bounded in-flight/retry capacity. They are safety limits, not tested
throughput commitments. See configuration. SQL writers are
serialized; one combined owner is required per storage root. Advisory local
locking is not a distributed/NFS coordinator.
Compatibility evidence#
| Area | Source of support/configuration | Evidence needed |
|---|---|---|
| Python | pyproject.toml; CI matrix 3.11–3.13 | Successful jobs for the target commit |
| OS | CI uses Ubuntu; no full OS support matrix declared | Target-platform validation |
| Dependencies | Ranges in pyproject.toml, exact graph in uv.lock | Locked install and compatibility check |
| Storage | SQLite FTS5, LanceDB, Kùzu | Native-store contracts plus representative capacity/recovery runs |
| Models | Explicit provider/model identity | Real-provider behavior, quality and cost measurements |
| Docker | Dockerfile and two Compose topologies | Image and deployment evidence for the target commit |
Unsupported or constrained#
Bitemporal queries, associative PPR, human legal certification, distributed
writers/hosted storage, Ladybug backend, online or tenant/dataset-scoped rebuild,
automatic retained-generation cleanup and guaranteed citation correctness are
not current Core guarantees. V4 retrieval performs no answer generation.
Adaptive planning is opt-in, bounded and not recursive.
