Foundgine
PostgreSQL 17 E2E
The repository has a real PostgreSQL 17 integration path.
The normal test suite does not require a database. The PostgreSQL tests run when FOUNDGINE_POSTGRES_CONNECTION_STRING is set.
CI always supplies that connection string.
PostgreSQL retrieval in the canonical lifecycle
PostgreSQL retrieval is a branch inside the canonical semantic lifecycle, not a second pipeline:
No retrieval strategy grants authority or bypasses semantic authorization.
Start PostgreSQL
docker compose -f docker-compose.postgres.yml up -d
Check:
docker compose -f docker-compose.postgres.yml ps
Check the version:
docker compose -f docker-compose.postgres.yml exec -T postgres \
psql -U foundgine -d foundgine_e2e -c 'select version();'
The container uses:
image: pgvector/pgvector:pg17
host port: 55432
database: foundgine_e2e
user: foundgine
password: foundgine
pgvector/pgvector:pg17 is a drop-in-compatible build of the official postgres:17 image with the pgvector extension preinstalled. It is required by Foundgine.Providers.Storage.PostgresVector’s E2E tests (PostgresVectorE2ETests), which call CREATE EXTENSION vector — something a vanilla postgres:17 image cannot satisfy.
The database uses a temporary filesystem, so it is safe to throw away after the test.
Run the tests
PowerShell:
$env:FOUNDGINE_POSTGRES_CONNECTION_STRING="Host=localhost;Port=55432;Database=foundgine_e2e;Username=foundgine;Password=foundgine"
dotnet test .\src\csharp\tests\Foundgine.E2E.Tests\Foundgine.E2E.Tests.csproj --configuration Release --filter "FullyQualifiedName~Postgres"
Bash:
export FOUNDGINE_POSTGRES_CONNECTION_STRING='Host=localhost;Port=55432;Database=foundgine_e2e;Username=foundgine;Password=foundgine'
dotnet test src/csharp/tests/Foundgine.E2E.Tests/Foundgine.E2E.Tests.csproj --configuration Release --filter "FullyQualifiedName~Postgres"
One command:
bash ./scripts/run-postgres-e2e.sh
PowerShell:
.\scripts\run-postgres-e2e.ps1
What the PostgreSQL tests prove
The PostgreSQL tests are the physical database proof for:
- generated identity correlation;
- dependency levels;
- compiler-owned correlation;
- batched mutation execution;
- real PostgreSQL execution;
- the pgvector-backed lexical grounding pipeline — semantic contract → lexicon projection → embedding → pgvector index → nearest-neighbor retrieval → graph-constrained resolution (
PostgresVectorE2ETests).
The tests that need a database are deliberately skipped when no connection string exists. This keeps ordinary local tests easy to run.
The dedicated PR job does not skip them.
Clean up
docker compose -f docker-compose.postgres.yml down --volumes --remove-orphans
Measurement gate
Do not optimize the PostgreSQL compiler before collecting real PostgreSQL measurements.
The target matrix is:
batch size: 1 / 10 / 50 / 500
depth: 1 / 2 / 3
Collect:
planning time
execution time
shared buffer hit/read/write
temporary read/write
WAL bytes
join type
sorts
materialization
estimated rows
actual rows
actual loops
The goal is to find the expensive part from the PostgreSQL plan instead of guessing.
Next: Testing