Compatibility and validation

The SQL dialect target is PostgreSQL 18 compatibility through ADBC (#118). The differential harness now selects a pinned PostgreSQL 18 reference; its existing curated upstream inputs retain PostgreSQL 16 provenance. The PostgreSQL 16 measurements below are historical, not PostgreSQL 18 evidence or a SQL:2023 Core conformance score. See the initial SQL surface audit and evidence requirements. The 2026-11-10 curated comparison at implementation revision b35a4739 passed 574/651 scored cases (88.2%), with 77 failures, two allowed extensions, two explicit exclusions and zero infrastructure errors. Of #359's 188 retained baseline failures, 122 now pass and 66 have explicit deferred dispositions; the other eleven failures are the original #360 storage-model cases. Deferred cases remain failing in the runner. The proposed query-profile milestone and categorized failure inventory explain a narrower future claim; no complete profile pass has been established.

On 2026-10-08, the validation suite against the default local Redis Software database (Redis 8.6.2, Search 8.6.11) had 360 passed, 8 skipped, 3 subtests passed, and 0 failed. The full Go suite also passed, with explicit skips for protocol-level ACL administration and resettable command counters that Software blocks. Those checks passed separately on the OSS three-shard compatibility environment. See Building and testing. The other validation results below are from earlier runs.

Server Version Connection
Redis Software (default), redislabs/redis:latest Redis 8.8.0, Search 8.8.8 (2026-11-04) redis://
Redis Open Source (redis:8.6.2 Docker image) 8.6.2, Search 8.6.0 redis://
Redis Open Source cluster, 3 shards (OSS Cluster API) 8.6.2, Search 8.6.0 redis://
Redis Cloud, single shard 8.6.2, Search 8.6.10 redis:// and TLS (rediss://)*†
Redis Cloud Pro, 2 shards, through the proxy endpoint 8.6.2 redis://†
Redis Cloud Pro, 2 shards, OSS Cluster API enabled 8.6.2 redis://†

Redis Software with verified 4× Query Performance Factor (QPF) is the default development, local test, and CI environment. All Redis Software test fixtures must enable 4× QPF before tests run. make redis-up pulls the redislabs/redis:latest image, bootstraps a single-node cluster, and creates a Redis 8.8 RAM database with Search and 4× QPF. Setup verifies six Search workers, mnp routing, and 32 connections through the management API, including on reused databases. Open Source standalone and cluster environments are optional compatibility targets; CI also tests the three-shard OSS Cluster API.

Earlier local and cluster runs had 325 passed, 0 failed. The Cloud runs marked † had 322 passed, 0 failed, before the two temporary-table ingest tests were enabled; they have not been re-run since. The TLS run marked * predates cluster support and has not been re-run since.

Those historical runs skipped statistics, unsupported constraints, a second catalog and transactions. Current source validation also exercises approximate row-count statistics. CHECK and PRIMARY KEY discovery run; UNIQUE and FOREIGN KEY discovery remain skipped.

Redis Flex (RAM + SSD) databases are not supported yet. Older builds are refused on connection; the tested 8.2.6-25 build fails during table creation. See Server requirements.

The differential harness compares rows, physical Arrow result types and SQLSTATEs with digest-pinned amd64 Postgres 16.15 and C collation (byte order). Run REDIS_URI=redis://localhost:6379/0 make -C go pgdiff. On October 9, 2026, its expanded curated corpus passes 327/603 scored cases (54.2%), including 13/13 smoke cases, 19/23 captured dbt cases, and 229/371 Postgres regression cases. Four more cases are separately classified as documented extensions or explicit scope exclusions, for 607 cases total. This measures the checked-in corpus rather than full Postgres conformance. Unresolved differences remain visible; fixture/infrastructure failures have a separate exit status. See the harness guide for the coverage inventories, pinned sources, type-comparison limits and manual CI run.

Latest Software retest (2026-11-04)

New RAM databases default to Redis 8.8; SOFTWARE_REDIS_VERSION overrides the requested version at creation. Both Software Compose services pull the latest image on startup. Existing databases retain their database version when reused. The default TLS, authentication, three-shard topology, and verified 4x QPF configuration are preserved.

The earlier checkout at 95d6583 was tested against Software 8.2.6-25 / Redis 8.8.0 / Search 8.8.8 using an isolated single-shard RAM database. go test -count=1 ./... reported one failure in TestTimeTZSessionZoneColumns; the other packages passed. The same focused test failed on Software 8.2.0-78 / Redis 8.6.2. Its expected row set assumes New York has a different UTC offset from January, which is false on this date. This is historical evidence for that revision and fixture, not a passing validation of the current default topology.

The Flex retest records improved Search capabilities and the remaining driver incompatibilities in issue #314.

The PR fixture check on current public main passed the focused TestSQLAliasesInExpressions test with three shards, TLS/authentication, and verified active 4x QPF (six workers, mnp, 32 connections). Setup now requests routing and connection count explicitly because Redis 8.8 can retain cmp/five connections when only the QPF factor is set.

Redis 8.8 defaults versus our test fixture

On Software 8.2.6-25, a fresh Redis 8.8 RAM database created with active 4x QPF returned the values below before explicit routing overrides. “Observed” describes that database, not every Redis Software deployment or an unconfigured database. In particular, we requested 4x QPF at creation; this does not establish the product's default QPF factor.

Setting Observed before routing overrides Required by our fixture
QPF active / scaling factor true / 4 (explicitly requested) true / 4
Effective Search workers 6 6
sched_policy cmp: choose the busiest eligible connection below its pipeline limit mnp: choose the connection with the fewest pending commands
conns 5 32

These are connections between the Software proxy and the shards, rather than application-side connection-pool settings. Redis documents cmp as “Closest to max_pipelined” and mnp as “Minimal pipeline” in the BDB API reference. For example, if eligible connections have 8, 3, and 0 pending commands, cmp chooses the queue of 8 if it is below the limit; mnp chooses 0.

The fixture preserves its established configuration to spread requests across more internal connections and make concurrency available to the Search workers. QPF controls Search capacity separately from the proxy's connection scheduling. Thirty-two connections do not mean 32 Search workers or CPU cores, and this check is not a benchmark showing mnp is always faster than cmp.

Setup explicitly applies the required routing and connection count when any effective QPF requirement is unmet, then waits for the management API to confirm all five values. It leaves an already compliant database alone and refuses to run tests if verification fails. TLS, authentication, and the default three-shard topology are fixture requirements preserved by this change; they were not inferred from the routing-default observation.

Restricted CONFIG GET responses in newer Software

The TLS/RBAC smoke check previously required an empty CONFIG GET reply for a restricted worker. Software 8.2.6-25 / Redis 8.8 instead returned NOPERM in both Linux architecture jobs. The check now accepts either an empty filtered reply or an explicit permission denial, records which one occurred, and still rejects any readable configuration. It also retains the separate checks that the worker cannot mutate configuration or write data. Unrelated server and connection errors are not treated as permission denials.