Contributes to Flagsmith/flagsmith#5663.

Initial scaffold for flagsmith-sql-flag-engine — sibling to flagsmith-flag-engine (Python) and flagsmith-rust-flag-engine (Rust). Translates SegmentContext predicate trees into SQL WHERE expressions.

The translator emits a predicate that runs against a single IDENTITIES table aliased i; trait-bound conditions become path-extractable lookups via the active Dialect. Snowflake reads from a VARIANT column, but the translator never assumes that storage shape. PERCENTAGE_SPLIT and :semver-marked comparators compile to inline pure-SQL.

Public API

from flag_engine.context.types import EvaluationContext, SegmentContext

from flagsmith_sql_flag_engine import TranslateContext, translate_segment
from flagsmith_sql_flag_engine.dialects import SnowflakeDialect

eval_context: EvaluationContext = {
    "environment": {"key": "n9fbf...3ngWhb", "name": "Production"},
}
ctx = TranslateContext(evaluation_context=eval_context, dialect=SnowflakeDialect())

segment: SegmentContext = ...  # your segment definition
where_expr = translate_segment(segment, ctx)

EnvironmentContext.key is used directly as the environment_id literal in emitted SQL — no separate integer PK. IDENTITIES holds a STRING environment_id column matching the engine's vocabulary.

What's included

• uv-managed Python package, src/ layout, py.typed
• Dialect protocol with SnowflakeDialect implementation; regex-flavour gating lives on the dialect via regex_supports, so future engines aren't forced into RE2's restricted feature set
• engine-test-data as a git submodule pinned to v3.7.0
• 55 unit tests (translator SQL-shape regression, no Snowflake required)
• 599 + 3 xfailed engine-parity tests parametrised over engine-test-data and a per-dialect DialectTestHarness; new dialects extend by writing one harness module under tests/harnesses/
• mypy strict across src/ and tests/
• CI workflow: lint + typecheck + unit + engine-parity gated on SNOWFLAKE_* secrets; 100% line + branch coverage gate via Cobertura
• Renovate config: CVEs only for deps, aggressive bumps on engine-test-data releases via the git-submodules manager (mirrors the Python engine's setup)
• CODEOWNERS: * @flagsmith/flagsmith-back-end, mirroring the Python engine

Operator coverage

Performance

Measured on COMPUTE_WH X-Small, query result cache disabled, warehouse SUSPENDed before each timed query so reads come off remote storage rather than warehouse-local SSD:

The %Split row above hashes on the implicit $.identity.key subject and so doesn't read traits; a PERCENTAGE_SPLIT against a trait property pays the per-row trait-read cost on top of the hash. Warehouse-scaling is roughly linear in warehouse size — divide the X-Small numbers by the warehouse-size multiplier for an estimate.