source-snowflake-e2e-tests

# source-snowflake-e2e-tests Remote end-to-end harness for `source-snowflake`. The shared `db-harness-lib` orchestration runs protocol commands against a per-run schema in the shared Snowflake integration account. ## Prerequisites - Docker, [`uv`](https://docs.astral.sh/uv/), and `jq`. - `gcloud` authenticated with read access to the GSM secret, or an existing `SNOWFLAKE_CONFIG_FILE`. - A clone of `airbytehq/airbyte`. ## Layout ``` source-snowflake-e2e-tests/ ├── SKILL.md ├── ../../../poe_tasks.toml (connector-root entrypoint) ├── scripts/ │ ├── run.sh, sf.py │ ├── fetch-config.sh, render-config.sh │ ├── start-backend.sh, reset-databases.sh, stop-backend.sh │ ├── apply-sql.sh, sweep-orphans.sh ├── fixtures/configs/base.template.json, utc-session.template.json ├── fixtures/sql/00-cursor-canary-base.sql ├── fixtures/sql/01-cursor-canary-boundary-row.sql └── cases/cursor-upper-bound.sh ``` ## Isolation model - Every run uses `PROVEFIX_<run_id>` and asserts that the schema is absent before both control and target runs. - A marker owned by the run permits reuse of the same schema between canary phases; unmarked pre-existing schemas are refused. - Control and target use the same schema with `--reset=fixture`; remote `--reset=backend` is rejected. - Teardown drops the schema in the run's exit trap. - `scripts/sweep-orphans.sh` lists or drops old `PROVEFIX_*` schemas, defaulting to four hours. The default run id is `YYYYMMDDhhmmss_<six-digit-host-checksum>_<pid>`, producing a schema-safe `PROVEFIX_<run_id>` name. The schema is owned by the scripts, while `sf.py` is the only code that talks to Snowflake. The default config comes from GSM secret `SECRET_SOURCE-SNOWFLAKE_KEY_PAIR__CREDS` in project `dataline-integration-testing`; fetched configs are marked with a sidecar and only those configs are removed during teardown, so caller-supplied configs survive. Set `SNOWFLAKE_CONFIG_FILE` to reuse a caller-supplied config. ## Cursor upper-bound canary The canary proves the cursor precision regression from airbytehq/airbyte#82705: ```bash cases/cursor-upper-bound.sh ``` It compares fixed `1.1.1` with control `1.1.0` in two phases. Phase 1 runs the initial sync on the control image with exact-microsecond base rows, producing the checkpoint a user on the bad version would have. Phase 2 inserts a sub-microsecond boundary row and replays that state against both versions: 1.1.0 loses the row because it floors the upper bound, while 1.1.1 emits it. The harness reuses the `INTEGRATION_TEST_WAREHOUSE_DESTINATION` X-Small warehouse shared with destination-snowflake integration tests. Never touch schemas outside `PROVEFIX_*`. ## Gotchas Snowflake uppercases identifiers, so use `--cursor-field=UPDATED_AT --streams=CURSOR_CANARY`. Rendered configs contain the private key; treat `REPRO_OUT` as 0600 material and never attach it. TIMESTAMP_NTZ cursor bounds are bound as TIMESTAMP_LTZ by the JDBC driver and compared through the session TIMEZONE; with the account default (America/Los_Angeles), narrow incremental windows match nothing. The harness sets `jdbc_url_params=TIMEZONE=UTC` so the canary isolates #82705; this session-timezone behavior appears to be a separate connector bug worth its own issue. The `--catalog` CDC guard (#85752) never fires here: source-snowflake has no CDC replication method.