Meridian Storage Query

meridian-storage-query is the engine-neutral Meridian V1 query library. It provides an immutable query AST, a mapping-first fluent API, canonical serialized Operations, capability inference, compile/startup validation, single-Binding planning, adapter translation contracts, and signed live-keyset cursors.

The library does not define a query Catalog. Consumers use a registered Catalog such as structured or evidence; deployment code selects Bindings and adapters. V1 has no NativeQuery, native-expression method, Engine selector, endpoint, or backend DSL escape hatch.

Install

python -m pip install meridian-storage-query==1.0.2

Python 3.12–3.14 is supported. Version 1.0.2 consumes the released meridian-storage-core==1.0.1 and meridian-storage-semantics==2.0.0 contracts.

Mapping-first query

from meridian_storage.query import distance_within, field, point, query

plan = (
    query("accounts.users")
    .where({"status": "active", "age": {"$gte": 18}})
    .select("id", "display_name")
    .order_by("display_name", "id")
    .page(size=50)
    .build()
)

assert plan.format_version == "meridian.operation.query.v1"

Specialized clauses use the same immutable builder:

search = (
    query("knowledge.documents")
    .where({"visibility": "public"})
    .full_text(
        "storage semantics",
        fields=("title", "body"),
        analyzer="icu",
        highlights=("body",),
        ranking="bm25",
        facets=("category",),
    )
    .build()
)

nearby = query("places.locations").where(
    distance_within(field("location"), point(-122.4, 37.8), 1_000)
).build()

For spatial predicates, use point, distance, or distance_within; for time-series windows, use time_range. Aggregation uses group_by and measure. traverse accepts either an explicit closed relation-Collection set or all_neighbors=True, which must be resolved against a captured registry fingerprint before execution.

Public pipeline

Catalog Expression
  -> normalized QueryOperation
  -> Schema/Resource/Binding validation
  -> inferred RequirementGraph
  -> NATIVE | EXACT_REWRITE | BOUNDED_RESIDUAL | REJECTED
  -> adapter-owned parameterized compilation
  -> normalized Data and signed cursor

validate_compile resolves exact Schemas, rejects cross-Binding plans, injects a deterministic identity tie-breaker, checks logical types and semantic profiles, and enforces a registered workload budget. validate_startup compares declared deployment requirements with authenticated Core capability manifests. QueryPlanner never silently broadens an unsupported query into a scan: rewrites require a registered equivalence and residuals require explicit row/byte bounds plus a proof function.

Adapter repositories implement QueryTranslator. Its input is a validated PlannedQuery and a logical TranslationContext; connection and Engine state remain private to the adapter. enforce_result_budget applies the validated value/byte limits after result normalization.

Pagination and diagnostics

CursorSigner issues opaque HMAC-SHA256 cursors bound to the plan, exact Schema fingerprints, registry fingerprint when applicable, scope fingerprint, unique sort tuple, page size, issuance, and expiry. Key rotation accepts old verification keys while issuing with one active key.

Diagnostics use stable MERIDIAN_* codes and logical references. Explain records redact secrets, credentials, endpoints, native statements, and backend DSL details.

Contracts and compatibility

• Operation, capability, and cursor contracts
• Public API manifest
• Architecture
• Adapter conformance
• Compatibility ledger
• Release procedure

The implementation is pinned to Meridian HLD revision 56 and Catalogs/Public Interfaces revision 70, with Query LLD revision 46 and Adapter LLD revision 24 as package-level authorities.

License

Apache License 2.0. See LICENSE and NOTICE.