faucet-source-singer

Support tier: Tier-2 / experimental. Best-effort — correctness bugs are fixed, but breadth of testing and upstream-drift tracking are not guaranteed.

A Singer tap bridge source for faucet-stream. It runs an existing Singer tap executable and adapts its stdout message stream into faucet records — so any of the hundreds of community taps can feed a faucet pipeline. Its counterpart, faucet-sink-singer, runs a Singer target as a faucet sink; the protocol types both share live in faucet-common-singer.

Honest trade-offs

• It reintroduces a runtime dependency. Most Singer taps are Python; a pipeline using this source needs that interpreter (and the tap) installed on the box. This is the one place faucet steps outside its "single static binary" story — use a native faucet source when one exists.
• Throughput is Singer-class, not faucet-class. Records cross a process boundary as newline-delimited JSON; expect tap-bound throughput, not faucet's native streaming numbers.
• Resume granularity depends on the tap. faucet checkpoints at the tap's own STATE messages; how coarse or fine that is (and whether re-emitted rows overlap) is a property of the individual tap. Pair with an idempotent sink for clean effectively-once (idempotent at-least-once) behavior.
• Stopping on Windows. On Unix the tap gets SIGTERM and a grace period before it is killed. Windows has no SIGTERM, so the tap gets the same grace period to exit on its own and is then terminated. executable must be something Windows can launch directly: an .exe (pip's console-script launchers are) or a .cmd/.bat named with its extension.

v0 scope

• Single-stream. Exactly the configured stream is emitted; RECORD messages for other streams are ignored. Multi-stream fan-out is future work.
• Handles RECORD, SCHEMA (pass-through — faucet sinks infer schema from records), and STATE (resume bookmark). ACTIVATE_VERSION / BATCH are logged and skipped.

Configuration

Environment — inherit_env

The tap runs as a child of faucet and, by default, sees faucet's whole environment, as it would from a shell. That includes whatever credentials faucet itself holds there (cloud keys, vault tokens, a faucet serve server's secrets). To keep them away from the tap:

inherit_env: false                 # only PATH, HOME, LANG, LC_ALL, TMPDIR
inherit_env: [TAP_API_TOKEN]       # that baseline plus the listed variables

Pass credentials the tap needs through tap_config (resolved by faucet and written to a private file) rather than through the environment.

Catalog-driven stream selection

Most database and Meltano-SDK taps sync nothing unless the target stream is marked selected in the catalog — a catalog passed through verbatim is a silent no-op. Parent-keyed taps go further: tap-github's issues stream only syncs when its parent repositories stream is also selected, even though faucet emits just issues.

faucet init --source singer --discover --executable <tap> --stream <name> runs the tap's discovery and writes a catalog with <name> — and any inferable parent streams — marked selected (both the stream-level flag and the breadcrumb: [] metadata). When the tap doesn't express a parent relationship in its catalog, faucet init prints a warning listing the other streams so you can select a parent manually. Extraction may be multi-stream (parents are pulled to satisfy the child), but faucet still emits only the configured stream.

ACTIVATE_VERSION / FULL_TABLE + an append sink can accumulate duplicates. A tap doing full-table reloads re-emits every row each run; an append-only sink (jsonl/csv/stdout) keeps them all. Pair such taps with a keyed sink (write_mode: upsert, key: [...]) so re-emitted rows converge.

Example

version: 1
pipeline:
  source:
    type: singer
    config:
      executable: tap-github
      stream: issues
      tap_config:
        access_token: ${env:GITHUB_TOKEN}
        repository: faucet-hq/faucet-stream
  sink:
    type: file
    config:
      path: ./out/issues.jsonl
  state:
    type: file
    config:
      path: ./state

Capabilities

The crate exports DESCRIPTOR (RFC 0013). It declares no optional source capabilities: no exactly-once replay, discovery, sharding, columnar or native path. faucet list --json prints the declaration.

Timeouts, cancellation and retries

Every read of the tap's output waits through the pipeline's CallContext (#880): the run's cancel token or any caller deadline ends the wait and stops the tap (SIGTERM, then SIGKILL). idle_timeout_secs still bounds a tap that goes quiet. Spawning the tap (for a run and for discovery, faucet init --discover) does not block the runtime: a transient Text file busy retry waits through the same context.

License

Licensed under either of Apache-2.0 or MIT at your option.