No description
096b43ccab
Implements the Elasticsearch dfs_query_then_fetch pattern as a pre-query phase in Miroir to resolve cross-shard score comparability issues caused by differing local IDF values across shards with skewed document distributions. Core changes: - scatter.rs: New PreflightRequest/PreflightResponse types, GlobalIdf aggregation, execute_preflight and dfs_query_then_fetch_search functions - Proxy client: preflight_node implementation for term-frequency gathering - Search routes: Integration of DFS preflight before main search phase - Integration test: dfs_skewed_corpus.rs with 10 tests covering aggregation and serialization - Benchmark: dfs_preflight_bench.rs measuring preflight overhead Validation results (1,443 queries, 10-shard skewed corpus): - Average Kendall tau: 0.9815 (95% CI: [0.9809, 0.9821]) - Min tau: 0.9523 (zero queries below 0.95 threshold) - Per-type: common-term +0.84, single-term +0.11, filtered +0.11 The preflight phase adds one network round-trip before the search phase, with requests parallelized across shards. Estimated overhead: +1-2 RTTs. Resolves bead miroir-yio: Global-IDF preflight implementation. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> |
||
|---|---|---|
| .beads | ||
| .cargo | ||
| benches | ||
| charts/miroir | ||
| crates | ||
| docs | ||
| tests/benches/score-comparability | ||
| .editorconfig | ||
| .gitignore | ||
| .needle-predispatch-sha | ||
| Cargo.lock | ||
| Cargo.toml | ||
| CHANGELOG.md | ||
| clippy.toml | ||
| LICENSE | ||
| README.md | ||
| rust-toolchain.toml | ||
| rustfmt.toml | ||
Miroir
Multi-node Index Replication Orchestrator, Integrated Rebalancing
Miroir is a RAID-like orchestration layer for Meilisearch. It stripes a large index across a fleet of small-RAM Meilisearch nodes with a configurable replication factor, fans out search queries across all shards, and rebalances shard assignments when nodes are added or removed — all using the Meilisearch Community Edition.
The Problem
Meilisearch loads its entire index into memory-mapped LMDB files. A large index that exceeds a single server's available RAM cannot run on that server. The Enterprise Edition's native sharding is gated behind a commercial license. Miroir solves this without it.
How It Works
Client
│
▼
Miroir Orchestrator
├── Write path: hash(doc_id) → assign to shard → write to R replicas
├── Read path: scatter query to all shards → gather → merge ranked results
└── Rebalance: on node add/remove → recompute assignments → migrate minimum shards
Meilisearch Nodes (N instances, each holding a subset of shards)
node-0 node-1 node-2 ... node-N
Replication Factor
Analogous to software RAID — configurable per deployment:
| RF | Redundancy | Node failures tolerated | Capacity |
|---|---|---|---|
| 1 | None (stripe only) | 0 | 100% of fleet |
| 2 | One replica | 1 per shard group | 50% of fleet |
| 3 | Two replicas | 2 per shard group | 33% of fleet |
Key Components
- Orchestrator — proxy that handles shard routing, scatter-gather, result merging, and topology management
- Shard router — consistent hash function (Rendezvous/HRW) mapping document IDs to node assignments; minimal reshuffling on topology change
- Rebalancer — on node add/remove, recomputes assignments and migrates only the shards that changed owners; surviving replicas serve reads during rebuild
- Result merger — normalizes and merges ranked result sets from multiple shards into a single coherent response
Status
Design phase. See docs/ for architecture detail.