view and serve

view and serve are reusable definition resources. They do not declare their own input wiring; a pipeline supplies from.

A view materializes pipeline output into tapstate's managed state store. A serve surface delivers output to systems outside tapstate. A pipeline can use either one or both; declaring a view does not require a serve block.

Current preview boundary

The Quickstart uses an inline view to materialize MySQL and PostgreSQL data in the managed store. serve.sync delivers pipeline output to an external target and can coexist with a view.

The v1 Schema also accepts serve.query and serve.push. They are declaration shapes, not executable surfaces in the current preview. Use the CLI, REST API, or MCP for bounded state inspection; those read surfaces are not a stable application-facing data API.

kind: view

version: tapstate/v1
kind: view
id: customer-state
primary_key: customer_id
storage:
  warm:
    collection: customers
    indexes: [email]

Reference the definition from a pipeline:

view:
  use: customer-state
  from: shaped-customers

An inline view uses the same primary_key and storage fields and also requires id and from.

For runtime materialization, primary_key is required and must be one column that identifies the output row. The current preview materializes the warm database layer. The Schema also describes hot and cold, but this release refuses those tiers instead of materializing them.

Set optional execution on an inline view or the reusable View resource. The view.use reference does not accept an execution override. View sinks default to cluster-wide parallelism 4; see Node execution.

View schema compatibility

Views do not accept schema: in v0.6. Remove it from local definitions; validate and apply otherwise return dsl.unknown-field. Stored definitions are cleaned during upgrade. This does not remove connector config.schema settings.

Explicit View keys

An explicitly selected View key can use a unique identity supported by the discovered source model. It is not an arbitrary key override. Updates and deletes must supply that key in the before image; otherwise the write is refused with engine.view-key-missing-from-before-image.

kind: serve

The following resource shows the v1 Schema shape. In the current preview, use the sync block for external delivery; do not use query or push to configure a live endpoint or delivery path.

version: tapstate/v1
kind: serve
id: customer-outputs
sync:
  - id: customer-store
    source: warehouse
    write_mode: upsert
    ddl: fail
query:
  - type: rest
    backend: customer-store
push:
  - id: customer-events
    source: event-bus
    topic: customer-events

Reference the definition from a pipeline:

serve:
  use: customer-outputs
  from: customer-state

An inline serve block uses the same sync, query, and push arrays and requires from.

sync

Each element requires a target connection ID in source and can include:

• id;
• write_mode: upsert or append;
• ddl: apply, ignore, or fail;
• on_full_load: append (default), clear, or fail;
• target table rename rules;
• connector-owned options.
• node-level execution for parallelism and batching.

Sync sinks default to cluster-wide parallelism 4. Set execution on each sync element; it is not a setting on the containing serve block.

on_full_load controls existing target rows before a new full load. append keeps them, clear removes them, and fail refuses a nonempty target. Resume, recovery, and CDC-only runs do not clear target rows. The connector must expose the capabilities needed by the selected policy.

Before starting a pipeline that writes a View or uses serve.sync, discover the schema for each pipeline source. Applying the resources does not perform that discovery. If a source schema is missing, the start is refused before data-plane components start, the pipeline reaches FAILED, and its failure code is actuation.source-schema-not-discovered.

Discover the missing source schema and then start the pipeline again. This precondition applies to both View and sync outputs.

Choose a write mode

upsert is the default. It requires every selected source table to have a primary key in its discovered schema so tapstate can identify updates and deletes. A unique index does not satisfy this check, and a key declared only on the target does not replace the source key.

Use append only for insert-only delivery. It does not apply source updates or deletes as keyed changes. Although the v1 grammar accepts tables[].pk, the current runtime does not use that field as a primary-key override.

Rename target tables

rename changes the target table or collection name without changing the pipeline output fields:

serve:
  from: shaped-orders
  sync:
    - id: warehouse-write
      source: warehouse
      write_mode: upsert
      rename:
        map:
          ORDERS: orders_current
        case: lower
        prefix: odp_
        suffix: _v1

An exact entry in map has priority. For every other table, tapstate applies case first and then adds prefix and suffix. Accepted case values are upper, lower, camel, and pascal.

query

The Schema accepts type: rest, graphql, or mcp. Its optional backend field can name a sync element. The current preview does not execute serve.query; use the read surfaces described above for inspection.

push

The Schema accepts a target connection ID in source and can include id, topic, format, and connector-owned options. The current preview does not execute serve.push.