Limits
Data model
| Limit | Value |
|---|---|
| Namespace name | [a-z0-9][a-z0-9_-]{0,62} (1–63 characters) |
| Relationships between two nodes | One per (type, start node, end node); a second CREATE of the same triple replaces the first |
| Node id | A UUID, assigned at creation (or supplied as _id by the bulk API) |
| Writers per namespace | One at a time; any number of readers |
| Constraints and indexes | On node properties only |
| Integer values | 64-bit signed; arithmetic wraps on overflow |
| Temporal types | DATE and UTC DATETIME only; no durations, times, zones or points |
Per-node property limits
Checked when node data files are built (flush and compaction). Exceeding one fails the build with an explicit error rather than truncating data.
| Limit | Default | Variable |
|---|---|---|
| Distinct property names per data file | 65,536 | NAMIDB_NODE_PROPERTY_MAX_PROPERTIES |
| Property name length | 1,024 bytes | NAMIDB_NODE_PROPERTY_MAX_NAME_BYTES |
| One property value | 64 MiB | NAMIDB_NODE_PROPERTY_MAX_VALUE_BYTES |
| Nesting depth of a value | 32 | NAMIDB_NODE_PROPERTY_MAX_VALUE_DEPTH |
| Elements in one value | 1,000,000 | NAMIDB_NODE_PROPERTY_MAX_VALUE_ELEMENTS |
| Decoded property page | 65 MiB | NAMIDB_NODE_PROPERTY_MAX_PAGE_BYTES |
| Property catalog per file | 16 MiB | NAMIDB_NODE_PROPERTY_MAX_CATALOG_BYTES |
Query budgets (server)
| Limit | Default | Setting | Result when exceeded |
|---|---|---|---|
| Read statement time | 30 s | --query-timeout | 504 timeout |
| Write statement time | 60 s | --write-timeout | 504 timeout, nothing committed |
| Rows any read operator may materialise | 1,000,000 | --query-row-cap | 413 row_cap |
| Working memory of one statement | 50% of --memory-max-bytes (off without a ceiling) | --query-memory-max-bytes | 413 query_memory_limit |
| Working memory of all running statements | 75% of --memory-max-bytes | derived | 503 memory_pressure |
| Whole HTTP request | ≥ 120 s (derived) | --http-request-timeout | 408 |
| Wait for the writer lock | 30 s | --writer-lock-timeout | 503 |
| Wait for read admission | 5 s | NAMIDB_QUERY_ADMISSION_WAIT_MS | 503 |
| Concurrent reads | CPU cores (at least 4) | NAMIDB_MAX_CONCURRENT_QUERIES | queued, then 503 |
| Concurrent full scans | 4 | NAMIDB_MAX_CONCURRENT_SCANS | queued |
| Concurrent HTTP requests | 1,024 | NAMIDB_HTTP_MAX_CONCURRENCY | queued |
Results are materialised before they are returned, which is why the row cap
and the per-query memory budget exist: they stop a runaway cross product or a
huge UNWIND before it exhausts memory. Page large results with
ORDER BY … SKIP … LIMIT or a key range. See
Memory.
Every operator checks the deadline and the cancel flag as it works (since
3.3.0, cross products, joins, filters, projections, DISTINCT, grouping,
sorting and list expressions included), so a statement stops within
milliseconds of its budget or of POST /v0/admin/queries/:id/cancel. Avoid
unanchored cartesian products of large labels (MATCH (a:Big), (b:Big))
anyway: they are expensive and usually hit the row cap or the timeout.
The embedded Python client has no default budgets; pass timeout_ms and
max_rows per call.
Request and message sizes
| Limit | Default | Setting |
|---|---|---|
HTTP /v0/cypher body | 2 MiB | fixed |
| Bolt message before authentication | 64 KiB | fixed |
| Bolt message after authentication | 64 MiB | --bolt-max-message-bytes |
| Bolt connections | 1,024 | NAMIDB_BOLT_MAX_CONNECTIONS |
| Bolt transaction idle time | 30 s | --bolt-tx-timeout |
| Bolt transaction lifetime | 5 min | NAMIDB_BOLT_MAX_TX_LIFETIME |
Send large writes as several UNWIND $rows batches within these sizes.
Search
| Limit | Default | Setting |
|---|---|---|
| Unbounded full-text result size | 64 MiB | NAMIDB_SEARCH_MAX_RESULT_BYTES (413 search_result_limit) |
| Document size for exact BM25 | 1 MiB | NAMIDB_BM25_MAX_DOCUMENT_BYTES (413 search_document_limit) |
| Transient search memory | 256 MiB | NAMIDB_SEARCH_WORKSPACE_MAX_BYTES (503 search_workspace_capacity) |
| Distinct string/boolean values per filter property inside a vector index | 4,096 | fixed; above it the property is filtered by the slower exact path |
| Filter postings per vector index body | 64 MiB | fixed; same fallback |
| Sidecar candidates for a filter the index cannot apply | 8,192 | NAMIDB_VECTOR_FILTER_ID_CANDIDATE_CAP (0 disables) |
| Filtered hybrid BM25 candidates | 65,536 | NAMIDB_HYBRID_TEXT_FILTER_CANDIDATE_CAP |
| Live segments per search index | 8 (2–32) | NAMIDB_SEARCH_LSM_MAX_SEGMENTS |
int8 quantization | METRIC cosine only | — |
Hitting a filter or candidate cap changes cost, not results: the query falls back to an exact path.
Multi-tenant capacity
| Limit | Default | Setting |
|---|---|---|
| Open namespaces | 100 (oldest idle evicted) | --max-namespaces (0 = unlimited) |
| Idle time before eviction | 1 h | --namespace-idle-timeout |
| Concurrent reads per namespace | off | NAMIDB_MAX_CONCURRENT_QUERIES_PER_NAMESPACE |
| Concurrent backups / restores per process | 1 (waiters time out after 30 s) | fixed |
Throughput
All writes to a namespace go through its single writer, and each commit is a
round trip to the store. On an object store, interactive write throughput is
therefore bounded by commit latency; batch writes with UNWIND, use the
Python bulk API, or enable --group-commit-window to share commits between
concurrent writers. Reads scale with cores and are not blocked by writes.
Spread independent graphs over namespaces to scale writes further.