alivirgo/major-ai-skills/elasticsearch

Elasticsearch Search Cluster AI Skill Guide

Overview & Engine Architecture

Elasticsearch is a distributed search and analytics engine: documents in indices (sharded, replicated), mappings define field types, queries use JSON DSL. Ingest is bulk-oriented; coordinators merge shard results. Mapping changes are mostly immutable—plan reindex + alias swap. OpenSearch shares many APIs; verify version-specific behavior.

Ingest (bulk / Beats / OTel)
    -> coordinating node
        -> primary/replica shards (Lucene segments)
        -> merge / refresh / flush
        -> ILM: hot -> warm -> cold -> delete
Visibility index (async) != transactional DB

When to use / when not to

Use when

• Full-text search, faceted browse, log/metrics search at scale
• Time-series indices with ILM rollover and retention
• Aggregations over denormalized documents

Do not use when

• You need primary OLTP with row-level updates as the main pattern (@postgresql)
• Join-heavy relational models without denormalization
• Someone wants dynamic mapping on unbounded user keys in production—block or constrain

Operational Capabilities & Agent Directives

1. Connection/auth safety: HTTPS + API keys or native security; never expose :9200 anonymously on the public internet; rotate keys; use least-privilege roles per index pattern.
2. Mappings: Explicit properties for prod fields; disable or limit dynamic mapping; use keyword for filter/sort/agg, text for search (multi-fields when both needed).
3. Aliases: Clients read/write `logs-write` / `logs-read` aliases; reindex to logs-v2, atomic alias flip—never delete old index until verified.
4. Backpressure: Bulk with bounded batch bytes/docs; exponential backoff on 429 and circuit_breaking_exception; watch thread pools and indexing pressure.
5. Index design: Shard count is hard to change—start conservative; use ILM rollover on size/age; avoid oversharding small clusters.
6. Backup/restore: Snapshot repositories (S3/GCS); restore creates new indices—rename/alias carefully; snapshot restore is not a merge into live index names without downtime plan.
7. Exactly-once myth: Ingest pipelines may retry; use _id or external version for idempotent indexing; consumers of search results still see near-real-time delay (refresh_interval).

Production examples

Index template + alias

PUT _index_template/products
{
  "index_patterns": ["products-*"],
  "template": {
    "settings": { "number_of_shards": 1, "number_of_replicas": 1 },
    "mappings": {
      "dynamic": "strict",
      "properties": {
        "name": { "type": "text", "fields": { "keyword": { "type": "keyword" } } },
        "price": { "type": "scaled_float", "scaling_factor": 100 },
        "created_at": { "type": "date" }
      }
    }
  }
}
 
POST _aliases
{
  "actions": [
    { "add": { "index": "products-000001", "alias": "products-read" } },
    { "add": { "index": "products-000001", "alias": "products-write", "is_write_index": true } }
  ]
}

Reindex cutover

POST _reindex
{ "source": { "index": "products-000001" }, "dest": { "index": "products-000002" } }
 
POST _aliases
{
  "actions": [
    { "remove": { "index": "products-000001", "alias": "products-write" } },
    { "add": { "index": "products-000002", "alias": "products-write", "is_write_index": true } },
    { "add": { "index": "products-000002", "alias": "products-read" } }
  ]
}

Bulk ingest (bounded)

curl -s -H 'Content-Type: application/x-ndjson' -XPOST localhost:9200/_bulk \
  --data-binary @bulk.ndjson
# On 429: reduce batch size, increase backoff, check disk watermarks

Technical troubleshooting matrix

Best practices

• Test analyzers with _analyze before shipping multilingual search.
• Use data streams + ILM for logs where version supports it.
• Monitor JVM heap, GC, indexing rate, merge throttling, disk watermarks.
• Cross-cluster search/replication: document lag and auth separately.

Limitations

• Relevance tuning is iterative; agents propose experiments with metrics (nDCG, click-through), not magic weights.
• Major version upgrades need compatibility matrices and often reindex.
• Not a substitute for columnar OLAP at petabyte aggregations (@clickhouse).

Related skills

• @opentelemetry — trace/log pipelines into Elastic stack
• @kafka — ingest buffer before indexing
• @nginx-hardening — TLS reverse proxy for Kibana/ES

Agent Operational Directive

MANDATORY: Production mapping changes require reindex + alias plan. Never delete indices without confirming alias targets and retention. Bulk ingest must include backoff on 429. Do not disable security for convenience.

Source anchors (research)

• Elasticsearch: Index lifecycle management
• Elasticsearch: Reindex API
• Circuit breaker settings