版本说明

Authentication restructured into scramAuth 和 x509 blocks. The sync-source authentication config now uses explicit scramAuth (username/password) and x509 (certificate) sub-blocks rather than flat fields, with stronger validation at startup. Existing configurations must be migrated to the new structure. Introduced in v1.70.1.

X.509 certificate authentication for sync-source connections. mongot can now authenticate to mongod and mongos using TLS client certificates as a fully supported alternative to SCRAM. Configure syncSource.replicaSet.x509 or syncSource.router.x509 with your certificate key file and CA file. Introduced in v1.70.1.

Encrypted key file support for TLS. Both the sync-source TLS connection and the gRPC server TLS accept a certificateKeyFilePasswordFile so private keys no longer need to be stored unencrypted on disk. Introduced in v1.70.1.

新的 syncSource.replicationReader config block. Operators can now control which replica-set member mongot reads from during initial sync and change-stream replication. Accepts a readPreference value and optional tagSets for targeting members with specific replica-set tags. The deprecated replicaSet.readPreference field has been removed. Introduced in v0.66.0.

新的 advancedConfigs top-level config block. Optional performance-tuning parameters are now grouped under a single advancedConfigs: block with sub-blocks for indexing, querying, replication, autoEmbedding, and ftdc. FTDC and disk monitoring settings have moved here from the top level. Introduced in v1.70.1.

Sharded cluster support through syncSource.router. A router sub-block in syncSource enables connecting to a mongos router for sharded deployments. Each shard requires its own dedicated mongot instance, with replicaSet mapping the instance to its specific shard and router handling cluster-wide coordination. Available since v0.40.0.

Sorted indexes (index-sort). Search indexes can now be configured with a sort order on one or more fields. Queries with a $sort matching the index sort key skip the full-collection sort phase, significantly improving latency for sort-heavy workloads such as recency or price-sorted results. Introduced in v0.66.0.

Nested vector search. $vectorSearch now supports querying vector fields embedded inside nested document arrays. Define a nestedRoot path in the index definition to issue per-item similarity searches within arrays of sub-documents. Introduced in v1.70.1.

Vector stored source. Vector search indexes can store original vector values alongside index data. When a $vectorSearch query requests returnStoredSource, raw vector values are returned directly from the index, reducing latency by eliminating a second read from mongod. Introduced in v0.65.0.

Automatic ENN fallback for low-recall vector search. mongot now detects when approximate nearest-neighbor (ANN) search would return too few results due to high document-deletion rates and automatically falls back to exact nearest-neighbor (ENN) search, improving recall quality without requiring query changes. Introduced in v0.66.0.

Drill-sideways faceting. Facet queries use Lucene's drill-sideways algorithm, computing facet counts across all documents in a single pass. This enables "show counts as if this facet filter were not applied" behavior common in navigation UIs, without a second query. Introduced in v0.69.0.

Facet bucket limit raised to 10,000. The maximum number of buckets returned for number and date facets has been raised from 1,000 to 10,000, enabling finer-grained aggregations over large value ranges. Introduced in v1.70.1.

Aligned index definition syntax for self-managed mongot. The index definition JSON format accepted by self-managed mongot has been unified with Atlas. Index definitions can be used interchangeably without modification. Introduced in v1.70.1.

Auto-embedding with Voyage AI in self-managed mongot. Self-managed mongot now supports automated embedding generation. Configure the embedding: block with API key files and mongot automatically generates and maintains vector embeddings from your text fields as documents are inserted or updated. Introduced in v1.70.1.

HNSW parameters configurable for auto-embedding indexes. The m and efConstruction HNSW graph construction parameters can now be specified in auto-embedding index definitions, giving control over the accuracy/throughput trade-off of the approximate nearest-neighbor index. Introduced in v0.65.0.

Prometheus metrics endpoint for self-managed mongot. Self-managed mongot exposes a Prometheus-compatible /metrics endpoint (configurable with metrics.address, default localhost:9946) for monitoring index health, replication lag, query latency, and resource usage. Available since v0.40.0.

HTTP health and readiness endpoints. The /health endpoint (configurable with healthCheck.address, default localhost:8080) returns {"status":"SERVING"} (HTTP 200) when mongot is ready and {"status":"NOT_SERVING"} (HTTP 503) otherwise, suitable for container liveness and readiness probes. Available since v0.40.0.

$listSearchIndexes support for self-managed mongot. Self-managed mongot now supports running $listSearchIndexes on a collection to inspect the state, definition, and status of all search indexes, matching the behavior available in Atlas. Introduced in v1.70.1.

Container-aware memory metrics. System memory metrics now read from cgroup limits when running inside a container rather than reporting total host physical memory, giving accurate available-memory figures in Docker and Kubernetes environments. Introduced in v0.69.0.

Replication pause thresholds aligned with Atlas. The thresholds at which mongot pauses index replication to avoid overwhelming mongod are now consistent between Atlas and self-managed deployments. Introduced in v0.66.0.

Lucene upgraded to 10.1. The underlying search library was upgraded from Lucene 9.x to Lucene 10.1, bringing query performance improvements, improved segment merge behavior, and correctness fixes for compound queries, sorted indexes, and vector search. Introduced in v1.70.1.

Java runtime upgraded to Java 21. The mongot binary now ships with Java 21 (LTS), replacing Java 17. Java 21 delivers JIT improvements, better garbage collection, and a longer support lifecycle. Introduced in v0.69.0.

Docker base image upgraded to Amazon Linux 2023. The mongot Docker image now uses Amazon Linux 2023, replacing Amazon Linux 2, providing newer glibc, updated OpenSSL, and a longer support lifecycle. Introduced in v0.69.0.

Search indexes protected from deletion when router configuration is inconsistent with cluster topology. In a sharded cluster where mongot lacked router configuration, a search index could be deleted when catalog metadata moved to another shard. mongot now protects search indexes from deletion and preserves catalog metadata integrity in this scenario. Fixed in v0.69.0.

Startup crash when syncSource.router topology mismatched replicaSet. mongot could crash during bootstrap in a sharded cluster if the router topology was inconsistent with the configured replicaSet. This is now caught at startup and reported as a configuration error with a clear message. Fixed in v1.70.1.

Indexes that failed during initial sync are now automatically recovered. A search index that failed partway through initial sync would remain in a FAILED state permanently, requiring manual intervention. mongot now detects and retries failed initial syncs on restart. Fixed in v1.70.1.

Incorrect STEADY to INITIAL_SYNC status transition. A bug in the Lucene-layer status computation could incorrectly transition a fully-indexed search index from STEADY back to INITIAL_SYNC, triggering unnecessary re-indexing. Indexes now correctly remain in STEADY state. Fixed in v0.67.0.

Replication shutdown when sync-source host is unavailable. When all configured replica-set members are unreachable for an extended period, mongot now cleanly shuts down replication rather than spinning indefinitely, preventing resource exhaustion and surfacing the outage through health-check failures. Fixed in v0.67.0.

Stale system metrics. System metrics reported through the Prometheus endpoint could become stale after extended uptime. The metrics collection loop is now correctly reset on each scrape cycle. Fixed in v1.70.1.

Leading wildcard query regression (Lucene 10). A regression introduced during the Lucene 10 migration caused leading-wildcard queries to return incorrect or missing results. Resolved in Lucene 10.1. Fixed in v1.70.1.

Pure should -clause query regressions (Lucene 10). Queries composed entirely of should clauses returned incorrect scores or missed documents after the Lucene 10 upgrade. Resolved in Lucene 10.1. Fixed in v1.70.1.

Nested vector search score inflation with scoreMode: "avg". $vectorSearch with nested vectors and scoreMode: "avg" returned inflated ANN scores compared to ENN results. ANN and ENN scoring are now consistent. Fixed in v0.67.0.

Nested vector search with binary quantization returned incorrect results. $vectorSearch on nested vector fields using binary quantization could return incorrect results due to a bug in the child-filter path. Fixed in v0.69.0.

Drill-sideways token facets on empty fields threw an exception. Facet queries using the token field type on collections where no documents had indexed the field would throw an invalid dim exception. The empty case is now handled correctly. Fixed in v0.69.0.

Sparse facet double-counting. Iterating the key array of a sparse SSDV field could double-count or skip entries in facet results. Facet counts are now accurate for fields with sparse data. Fixed in v0.69.0.

Autocomplete operator dropped the final prefix token. When a user typed a word mid-completion, the autocomplete operator would sometimes discard the prefix token for the last word, returning results only for prior complete words. Fixed in v0.65.0.

Dropped collection metadata not cleaned up. When a MongoDB collection was dropped, stale index catalog metadata could accumulate and cause incorrect state if the collection was subsequently recreated. Metadata is now correctly removed on collection drop. Fixed in v1.70.1.

Native memory leak in FTDC data collection. A native memory leak in the FTDC collection routine caused mongot's off-heap memory footprint to grow over time. Fixed in v0.66.0.

Auto-embedding zombie lease race condition. A race between heartbeat checks and state-transition logic could leave an auto-embedding leader lease in a zombie state after the holder stopped. The lease state machine now uses atomic state checks to eliminate this window. Fixed in v0.69.0.

Materialized view full-replace incorrectly scoped for _id.<subfield> filters. Auto-embedding indexes whose filter included a sub-field of _id would incorrectly replace the entire materialized view on update, discarding out-of-filter documents. The update is now correctly scoped. Fixed in v1.70.1.

Security dependency upgrades. Third-party libraries with known CVEs were updated: Netty upgraded to 4.1.132+, Jackson upgraded to 2.18.6, Logback-core updated, Plexus-Utils upgraded to 4.0.3, and netty-tcnative upgraded to 2.0.79.Final. Applied in v0.69.0.