streamling source plugin — Agent Skill

A source is the head of a pipeline: it owns its schema, produces RecordBatches on demand, and drives checkpointing. Sources are pull-based — streamling calls generatebatch() repeatedly and stops when isrunning() is false or the pipeline drains.

Prerequisite: [streamling-plugin-basics](skill://streamling-plugin-basics) (crate setup, registration, lifecycle, errors).

The trait

#[async_trait]
pub trait SourcePlugin: SupportsGracefulShutdown + Send + Sync {
    async fn initialize(&self) -> Result<(), PluginError>;
    fn output_schema(&self) -> Result<SchemaRef, PluginError>;
    fn labels(&self) -> Vec<PluginLabel> { Vec::new() }            // optional identity
    async fn generate_batch(&self) -> Result<RecordBatch, PluginError>;
    async fn process_checkpoint_marker(&self, epoch: CheckpointEpoch) -> Result<(), PluginError>;
    async fn process_checkpoint_finalizer(&self, epoch: CheckpointEpoch) -> Result<(), PluginError>;
}

Constructor (called by registerpluginsource! — no schema arg, the source owns it):

pub fn new(
    rt: PluginAsyncRuntimeObj,
    state_backend_factory: PluginStateBackendFactory,
    metrics_recorder: PluginMetricsRecorder,
    options: HashMap<String, String>,
) -> Result<Self, PluginInitializationError>

Output schema and the gsop column

A source emits change-data with a per-row operation column. Import the canonical name and include it:

use streamling_plugin::api::STREAMLING_COLUMN_NAME_OP; // == "_gs_op"

let schema = Arc::new(Schema::new(vec![
    Field::new("id", DataType::Utf8, false),
    Field::new("value", DataType::Int64, false),
    Field::new(STREAMLING_COLUMN_NAME_OP, DataType::Utf8, false),  // "i" | "u" | "d"
]));

gsop values: "i" insert, "u" update, "d" delete. Downstream sinks use it to apply upsert/delete semantics. If your source is append-only, emit "i" for every row — but the column is still expected by the engine.

output_schema() must return the same schema for the life of the instance. Build it once in new() and clone it.

generate_batch — the pull contract

• Return a non-empty RecordBatch when you have data.
• Return an empty batch (RecordBatch::newempty(self.schema.clone())) when there is no data right now. The engine treats this as "not done, just idle" and will poll again. Do not block indefinitely inside generatebatch — sleep briefly then return empty if still waiting, so shutdown stays responsive.
• Return Err(PluginError::..) to fail the pipeline.

Building a batch with Arrow builders:

async fn generate_batch(&self) -> Result<RecordBatch, PluginError> {
    let rows = self.buffer.next_batch().await;
    if rows.is_empty() {
        return Ok(RecordBatch::new_empty(self.schema.clone()));
    }
    let mut ids = arrow::array::StringBuilder::new();
    let mut vals = arrow::array::Int64Builder::new();
    let mut ops = arrow::array::StringBuilder::new();
    for r in &rows {
        ids.append_value(&r.id);
        vals.append_value(r.value);
        ops.append_value("u");                 // or "i"/"d"
    }
    Ok(RecordBatch::try_new(self.schema.clone(), vec![
        Arc::new(ids.finish()), Arc::new(vals.finish()), Arc::new(ops.finish()),
    ]).map_err(|e| PluginError::ArrowError(e))?)
}

Fetcher / buffer / runner decomposition

Anything beyond a toy source should be split into three pieces (this is the shape every non-trivial production source converges on):

• Fetcher — pure async fetchpage(cursor) -> (Vec<Item>, nextcursor). No streamling types; trivially unit-testable against a mock HTTP client.
• Buffer — an ordered, bounded queue that the runner fills and generate_batch drains. Caps memory and provides backpressure.
• Runner (the SourcePlugin impl) — wires fetcher + buffer + checkpoint state to the streamling lifecycle: starts a background poll loop in initialize, drains in generatebatch, persists progress in processcheckpoint_finalizer, stops in terminate.

Keep the streamling trait impl thin; push logic into the fetcher/buffer so you can test it without the FFI.

Resumable pagination via the state backend

The durable-resume pattern: store a cursor/offset in PluginStateBackend, advance it only on processcheckpointfinalizer (the commit point), and restore it in initialize.

#[derive(Serialize, Deserialize, Clone, Debug)]
struct Cursor { last_ts_ms: i64, token: Option<String> }

pub struct ApiSource {
    schema: SchemaRef,
    client: ApiClient,
    buffer: Arc<OrderedBuffer<Row>>,        // shared with the poll loop
    state: Arc<PluginStateBackend<Cursor>>,
    running: Arc<AtomicBool>,
    cursor: OnceLock<Cursor>,               // set in initialize
}

#[async_trait]
impl SourcePlugin for ApiSource {
    async fn initialize(&self) -> Result<(), PluginError> {
        // Restore committed progress; fall back to the configured start.
        let mut cursor = self.state.get().await.map_err(PluginError::State)?
            .unwrap_or_else(|| Cursor { last_ts_ms: self.start_ts, token: None });
        // never regress past the configured start
        if cursor.last_ts_ms < self.start_ts { cursor.last_ts_ms = self.start_ts; }

        let buf = self.buffer.clone();
        let client = self.client.clone();
        let running = self.running.clone();
        // background poll loop (detached; Arcs keep it alive)
        tokio::spawn(async move {
            while running.load(Ordering::SeqCst) {
                let page = match client.fetch_page(&cursor).await {
                    Ok(p) => p,
                    Err(e) => { tracing::warn!(error=%e, "fetch failed; will retry"); continue; }
                };
                if page.items.is_empty() {
                    tokio::time::sleep(Duration::from_millis(60_000)).await;
                    continue;
                }
                cursor = page.next_cursor();          // monotonic advance
                buf.push(page.items).await;
            }
        });
        let _ = self.cursor.set(cursor);
        Ok(())
    }

    async fn generate_batch(&self) -> Result<RecordBatch, PluginError> {
        let rows = self.buffer.next_batch().await;    // drains up to batch_size
        if rows.is_empty() { return Ok(RecordBatch::new_empty(self.schema.clone())); }
        // advance an in-memory high-water mark (never regress)
        if let Some(last) = rows.last() {
            self.last_ts_ms.fetch_max(last.ts_ms, Ordering::AcqRel);
        }
        build_batch(&self.schema, &rows)
    }

    async fn process_checkpoint_marker(&self, _e: CheckpointEpoch) -> Result<(), PluginError> { Ok(()) }

    async fn process_checkpoint_finalizer(&self, _e: CheckpointEpoch) -> Result<(), PluginError> {
        // THE commit point: persist durable progress here, never in generate_batch.
        let cursor = Cursor {
            last_ts_ms: self.last_ts_ms.load(Ordering::Acquire),
            token: self.cursor.get().and_then(|c| c.token.clone()),
        };
        self.state.put(cursor).await.map_err(PluginError::State)
    }
}

Why finalizer, not process_batch: the finalizer fires only after the epoch is durably committed across the whole pipeline. Writing the cursor here gives exactly-once resume — a restart replays from the last committed cursor, never skipping data or emitting it twice durably.

Monotonic cursor: always fetch_max/max — never let a late or reordered page move the high-water mark backwards.

Two pagination flavors

• Numeric/height pagination (keyed by a monotonic u64, e.g. a sequence number or block height): use a sequential source — pages are height..height+N, dedup by height, resume stores the last height.
• Opaque cursor + timestamp (server-issued nextpage token plus a monotonic timestamp): use a cursor source — resume stores {lasttsms, token}, and on restart you re-request from the token or re-scan from lastts_ms. Catalog/reference APIs that change in place (a row's fields update over time) benefit from content-hash dedup: hash each row excluding its observed-at field, and only forward rows whose hash changed since last emit (in-process map, cleared on restart → one at-least-once re-emit, which is the documented tradeoff).

Snapshot-then-tail (hybrid) sources

When a dataset has a finite historical snapshot and then a live tail, run the snapshot to completion first, record the handoff point in state, then switch to polling the tail from that point. Keep both phases writing the same schema and gsop. The same source can expose both phases under one SourcePlugin; streamling also has a native hybrid source type if you'd rather compose two providers.

Backpressure

If your source can produce faster than the downstream can consume, bound the output channel:

register_plugin_source!("my_plugin", "api_source", ApiSource);
set_plugin_output_buffer!("my_plugin.api_source", 1);   // capacity 1: strict backpressure

A capacity of 1 makes the source block (cooperative, via the channel) until the downstream drains — useful for heavy transforms that can't keep up. Larger capacities decouple producer/consumer at the cost of memory. Pair with a bounded internal buffer so your poll loop naturally stalls instead of allocating unbounded memory.

Shutdown

#[async_trait]
impl SupportsGracefulShutdown for ApiSource {
    fn is_running(&self) -> bool { self.running.load(Ordering::SeqCst) }
    async fn terminate(&self) -> Result<(), PluginError> {
        self.running.store(false, Ordering::SeqCst);   // poll loop exits on next check
        Ok(())
    }
}

Keep is_running() returning true until initialize has run (e.g. gate on a OnceLock being set), so the engine doesn't bail before you've started.

Heavy row-building: use spawn_blocking

If converting fetched items into an Arrow batch is CPU-heavy (wide rows, many columns), move it off the async thread:

let schema = self.schema.clone();
let batch = tokio::task::spawn_blocking(move || build_batch(&schema, &items))
    .await
    .map_err(|e| PluginError::Internal(format!("blocking task: {e}")))??;

Common mistakes

• Blocking forever in generate_batch. Sleep-then-empty when idle, so terminate is prompt.
• Persisting cursor in processbatch. Breaks exactly-once. Use processcheckpoint_finalizer.
• Regressing the cursor. Use fetch_max/max on restore and on each batch.
• No gsop column. The engine and CDC-aware sinks expect it.
• Schema drift. output_schema() must be stable for the instance's lifetime.
• Unbounded buffer under backpressure. Cap the channel with setpluginoutput_buffer! and the internal buffer together.

Related skills

• [streamling-plugin-basics](skill://streamling-plugin-basics) — registration, lifecycle, options/secrets, state backend API.
• [streamling-sink-plugin](skill://streamling-sink-plugin) — the consuming half; sink checkpoint acks mirror source finalizers.