AgentSkillsCN

index-lifecycle

深入理解并调整Searchlite的存储/索引流水线（WAL → 分段构建 → 清单/压缩），同时探索各类存储后端；适用于处理数据摄入、持久性保障、快速字段优化，或分段布局的开发工作。

SKILL.md
--- frontmatter
name: index-lifecycle
description: Understanding and modifying Searchlite’s storage/indexing pipeline (WAL → segment build → manifest/compaction) and storage backends; use when working on ingest, durability, fast fields, or segment layouts.

Index Lifecycle Cheatsheet

The Searchlite mental model (one paragraph)

Searchlite serves one index directory per instance (CLI path or HTTP --index). A single writer buffers operations, appends them to a WAL, and on commit writes a new immutable segment and atomically updates a manifest. Readers consult the manifest to know which segments exist; compaction merges segments and drops deleted/obsolete data.

Glossary

  • •Schema: declares fields, analyzers, what’s stored vs fast vs indexed.
  • •WAL: write-ahead log for durability of add/delete operations.
  • •Segment: immutable set of postings/docstore/fast fields written on commit.
  • •Manifest: authoritative list of segments that define the current index state.
  • •Refresh: reload readers so they observe the latest manifest.
  • •Compaction: rewrite live docs into fewer segments; cleans duplicates/old versions.

Lifecycle steps (authoritative sequence)

1) Init

  • •Creates the index directory and writes schema/manifest state so it can accept docs.

2) Ingest (add/update/delete)

  • •Docs are upserted by primary key (doc_id_field, default _id)
  • •Deletes are queued by id
  • •Important: writes are buffered/queued; they are not visible to search yet

3) Commit

Commit is the “durable + visible” boundary:

  • •writer flushes buffered ops
  • •builds new segment files
  • •persists them
  • •atomically updates manifest

Durability contract:

  • •segment files/manifests are fsync’d on write
  • •WAL truncation only happens after manifest is persisted/synced
  • •manifest uses atomic rename + directory fsync

Crash window (expected behavior):

  • •if the process dies after manifest persisted but before WAL truncation, WAL replay re-applies the last batch
  • •no data loss; compaction cleans extra generations

4) Refresh (reader visibility)

  • •With HTTP, visibility may require explicit POST /refresh unless configured with --refresh-on-commit.

5) Search

Search uses:

  • •BM25 scoring by default (with pruning modes)
  • •phrase and fuzzy matching
  • •filters on fast fields
  • •aggregations on fast fields
  • •optional highlights, collapse/inner_hits, suggest, rescore, profile/explain

Execution modes (perf/correctness relevant):

  • •bm25 full evaluation
  • •wand exact pruning (default)
  • •bmw block-max WAND pruning (tunable block size)

6) Maintain

  • •Inspect: see manifest + segments
  • •Stats: see doc/segment counts
  • •Compact: merge segments and drop tombstoned/obsolete docs

File layout expectations (conceptual)

The exact file names are an implementation detail, but the directory contains:

  • •a manifest file (authoritative segment list)
  • •a WAL file
  • •per-segment files (postings/docstore/fast/meta/etc)
  • •optional vector index structures when vectors is enabled

When changing segment layout, preserve:

  • •compatibility with existing indexes or provide a migration story
  • •checksum/validation (if present)
  • •atomicity during commit and cleanup after rollback/failed commit

Schema invariants that drive the whole system

  • •doc_id_field is required and is the key for upsert/delete semantics.
  • •stored: true controls whether values can be returned/highlighted.
  • •fast: true controls filters + aggregations performance/feasibility.
  • •Nested fields flatten to dotted paths (e.g. comment.author) while preserving stored nested structure in responses.
  • •Nested filters must be expressed with Nested blocks so clauses bind to the same object instance.

Aggregations/filters coupling (easy to break)

  • •terms, significant_terms, rare_terms require a fast keyword field.
  • •numeric/date aggregations require fast numeric fields.
  • •multi-valued fields contribute multiple values to metrics (doc_count remains per-doc).

If you modify fast-field encoding, bucket iteration, or numeric parsing:

  • •add tests that validate aggs outputs (not just existence of buckets).

Vector search (feature vectors)

Vector search is approximate ANN (HNSW):

  • •dimension mismatches are rejected
  • •cosine vectors are normalized automatically
  • •tune recall/perf with candidate_size and ef_search
  • •vector_filter can reduce candidate selection set

If you touch vector integration:

  • •test both vector-only and hybrid search paths
  • •confirm response includes vector_score when vector search runs

“Where to make changes” guidance

  • •Core indexing/query/durability lives in searchlite-core
  • •CLI wiring + UX lives in searchlite-cli
  • •HTTP server wiring + limits + endpoints live in searchlite-http (and/or the CLI http subcommand)
  • •wasm storage backend + bindings live in searchlite-wasm
  • •C ABI lives in searchlite-ffi

Keep boundaries clean:

  • •searchlite-core should remain the source of truth for correctness; CLI/HTTP adapt surfaces around it.