Meridian OpenSearch Adapter

meridian-storage-opensearch is the independently released Meridian V1 adapter for eventually consistent structured Catalog search projections on OpenSearch. It is one Python distribution in one public Apache-2.0 repository.

Capabilities

The sole operation surface is meridian.structured.search@1.0.0. The adapter provides:

• closed, fingerprinted mappings for full-text and declared document-path projections;
• unrestricted BCP 47 language metadata with ICU fallback, with English and Chinese initial conformance;
• deterministic document identities, externally versioned upserts, tombstones, and bounded bulk partial outcomes;
• mapping-first Query V1 translation for the released logical operator allowlist;
• mandatory tenant/scope, Resource, and tombstone filters;
• signed search_after cursors, plus deployment-conditional point-in-time cursors;
• logical RecordRef results, bounded facets/highlights, and fail-closed partial-search handling;
• explicit generation create/verify/activate/rollback/retire hooks with atomic alias cutover; and
• authenticated engine, ICU plugin, topology, health, alias, analyzer, replica, and mapping probes.

The exact capability manifest, supported engine patch releases, and hard limits are available from adapter_descriptor(), capability_manifest(), and query_capabilities().

Install

python -m pip install meridian-storage-opensearch==1.0.0

The distribution requires the released Meridian Core, Query, and Semantics packages at exactly 1.0.0. Artifact hashes and locked design revisions are recorded in compatibility.json.

Configure through a Meridian Binding

Platform or Vangu IaC supplies the endpoint or service reference, secret references, compiled layouts, topology expectations, and exact capability fingerprint. Credentials never belong in an endpoint URL or settings document.

{
  "adapterId": "org.meridian.storage.opensearch",
  "engineProfile": "opensearch",
  "engineVersion": "2.19.1",
  "serviceRef": "platform://search/case-index",
  "physicalNamespace": "meridian",
  "requiredCapabilityFingerprint": "sha256:<64 hex characters>",
  "settings": {
    "indexPrefix": "meridian",
    "layouts": {
      "structured:investigation.articles": {
        "formatVersion": "meridian.opensearch.layout.v1"
      }
    },
    "requiredPlugins": ["analysis-icu"],
    "pitEnabled": true
  }
}

layouts must contain the complete value returned by MappingCompiler.compile(...).to_dict(); the abbreviated object above only illustrates placement. Use expected_capability_fingerprint(engine_version, settings) when producing the Binding pin.

Projection and search APIs

Adapter-internal projection workers use ProjectionExecutor with an IaC-selected write alias. Consumers do not receive this object and never see aliases, index names, mappings, analyzers, or OpenSearch DSL. Consumers submit released mapping-first Expressions or serialized Operations through Meridian Core.

The projection is always derived and eventually consistent. A successful authoritative write does not imply immediate search visibility. refresh=wait_for is available to controlled projection and conformance workflows; it is not an authoritative transaction guarantee.

Ownership boundary

This package owns translation, mappings, analysis definitions, indexing requests, cursor behavior, generation hooks, probes, and redacted engine-failure normalization. It does not create a Search or Projection Catalog and does not expose native queries, regex, wildcard/query-string syntax, scripts, credentials, endpoints, or physical lifecycle details.

Platform/Vangu IaC owns engine selection, provisioning/reference, state, identity and ACLs, migration orchestration, recovery, and lifecycle. See architecture, migrations, and the failure model.

Verify

python -m pytest -q tests/unit tests/contract tests/packaging
./scripts/run-single-conformance.sh
./scripts/run-cluster-conformance.sh
python -m build
python scripts/verify_artifacts.py dist

The Docker harness pins OpenSearch 2.19.1 by digest and installs its matching analysis-icu plugin. Cluster conformance uses three eligible/data nodes, one replica, green-health verification, and a deliberate node loss. See conformance evidence.

License and security

Licensed under the Apache License 2.0. See LICENSE, NOTICE, and SECURITY.md.