[rgamba]

What

A client-facing way to ask how many child workflows a parent has, optionally only those in a given set of statuses:

workflowsService.countChildWorkflows("parent-1")                          // any status
workflowsService.countChildWorkflows("parent-1", setOf(RUNNING, WAITING)) // only these

Design

• Lives on WorkflowsService, next to countWorkflowsByStatus. It reads the store directly, so its result changes between replays. The KDoc says to call it from client code or from inside an action, never directly in a workflow method.
• Two overloads instead of an optional set. Statuses are taken as a java.util.Set. The status overload rejects an empty set instead of reading it as "any status", so a caller whose computed set ends up empty gets an error rather than a count of every child.
• New WorkflowStore.countChildWorkflows, implemented for MySQL and SQLite as one COUNT(*) scoped to the owner. It is a default method that throws UnsupportedOperationException, as findWorkflows does, so WorkflowStore implementations outside this repo keep compiling. The testutils tracing store delegates to it.
• New index. parent_workflow_id had no index, so the count would have scanned every instance of the owner. This adds (owner, parent_workflow_id, status) through a MySQL migration (V7) and its SQLite counterpart (V3). It covers both query shapes, so the count never reads table rows.

Trade-offs

• The index adds a small write cost to every workflow insert and status update, in exchange for a bounded count per parent.
• On MySQL 5.6+ InnoDB, adding a secondary index is online and does not block writes. On a large workflow_instances table it still takes time and I/O, so deployments that apply migrations by hand should schedule V7 accordingly.

Testing

• Shared store test (BaseWorkflowStoreTest, runs against both MySQL and SQLite). It covers status filtering, multiple statuses, a sibling parent's children, top-level workflows, the same parent ID under another owner, and a parent that doesn't exist. A second shared test points a store at a table prefix with no tables and checks that the real SQLException comes back wrapped in InternalError. Both pass on SQLite (45/45 in SqliteWorkflowStoreTest). The MySQL run happens in CI; embedded MariaDB can't install on my machine.
• End-to-end (BaseWorkflowIntegTest). It starts two real child workflows through the factory and counts them through WorkflowsService: 2 in total, 2 COMPLETED, 0 RUNNING, and 0 for an unrelated parent. It passes.
• Unit tests (WorkflowsServiceTest) cover both overloads and the empty-set rejection. WorkflowStoreTest covers the interface default for stores that don't override it.
• Index use, SQLite EXPLAIN QUERY PLAN:

  SEARCH workflow_instances USING COVERING INDEX workflow_instance_index_by_parent (owner=? AND parent_workflow_id=?)
  SEARCH workflow_instances USING COVERING INDEX workflow_instance_index_by_parent (owner=? AND parent_workflow_id=? AND status=?)
  

• Migrations. The SQLite migration applies through Flyway in the file-backed factory test. The MySQL migration is applied by the MySQL suites in CI.
• Full local run across the root, testutils and state-machine projects: 1347 tests. The only failures are the three MySQL-backed suites, which fail during MariaDB4j setup on this machine. spotlessCheck is clean.

[rgamba]

statuses should be a set not a list

[rgamba]

Done in 3ed06fc — statuses is now a java.util.Set<WorkflowInstance.Status> (an empty set still means any status). I made the client-facing WorkflowsService.countChildWorkflows(parentWorkflowId, statusViews) overload take a Set too, so both layers agree; it still rejects an empty set.

• Ricardo's agent

[rgamba]

don't use vavr type here

[rgamba]

Done in 3ed06fc — both the MySQL and SQLite implementations (and the testutils tracing delegate) take a plain Kotlin Set now, with no vavr types in the signature. Placeholders are built with joinToString and bound with forEachIndexed.

• Ricardo's agent