fix(compose): default postgres volume to per-task (concurrency-safe)

[kentwelcome]

Problem

A full DataAgentBench run with concurrency.trials: 2 corrupts whole Postgres-backed datasets. Observed: all 13 crmarenapro cells errored with dependency failed to start: container …-dab-postgres-1 is unhealthy, and the compose stdout showed:

volume "dab-postgres-data-crmarenapro-v1" already exists but was created for
project "crmarenapro-q1__buzfvx7" (expected "crmarenapro-q1__v8mtc8f")

The 13 cells all started within ~65s of each other (concurrency), dropping the entire dataset from the board and forcing abstains on other PG-backed passers mid-run (could not translate host name).

Root cause

_postgres_volume_name(...) with the default postgres_volume_mode="reuse" keys the writable PGDATA named volume on (dataset_name, schema_version) only — e.g. dab-postgres-data-crmarenapro-v1 — shared across every query of a dataset (an intentional "run init.d once per dataset" optimization). Two failure modes:

1. Intra-run concurrency: under trials:2, two same-dataset cells mount the same writable PGDATA dir simultaneously. Postgres locks its data dir, so the second container never goes healthy → up --wait rc=1 → every cell of the dataset errors.
2. Cross-run persistence: the named volume survives across runs (keyed only on dataset+schema), so a prior run's project owns it → the "created for project X (expected Y)" warning and possibly stale data.

(dab-mongo is unaffected — it uses read-only bind mounts, no shared writable named volume.)

Fix

Flip the operator-facing default from reuse to fresh (per-task-unique volume, keyed on task_id) in cli.py and prepare_dataset_tasks(...). Each cell gets its own PGDATA volume (e.g. dab-postgres-data-crmarenapro-v1-crmarenapro-q7), eliminating both the intra-run collision and cross-run staleness. reuse stays available and documented for serial single-trial runs.

generate_compose's low-level default is intentionally left reuse so its existing API-documenting unit tests are untouched; the change is at the operator entry points.

Cost

Per-cell DB restore instead of once-per-dataset. Negligible: the worst concurrent case, crmarenapro (13 queries), is an 8.5 MB / 90k-line dump that restores in a few seconds — well inside the pg healthcheck budget (20×5s). Most datasets are far smaller.

Tests

Added test_same_dataset_tasks_get_distinct_postgres_volumes (red under the old default, green after). All 173 plugin unit tests pass (restart-policy, healthcheck, depends_on included).

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR changes the razorback-plugin-dab operator defaults to use per-task Postgres named volumes (instead of dataset-reused volumes) to avoid concurrency-related PGDATA directory collisions during concurrent DAB runs.

Changes:

• Flip the operator-facing default postgres_volume_mode from "reuse" to "fresh" in prepare_dataset_tasks(...).
• Flip the CLI default --postgres-volume-mode from "reuse" to "fresh" and strengthen the help text warning about concurrency.
• Add a unit test asserting that two tasks from the same dataset materialize with distinct Postgres PGDATA volume names under the operator default.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
packages/razorback-plugin-dab/tests/unit/test_compose_dataset_volume.py	Adds a regression test ensuring same-dataset tasks get distinct PGDATA volumes under the operator default.
packages/razorback-plugin-dab/src/razorback_plugin_dab/generate/prepare.py	Changes the default postgres_volume_mode to "fresh" and updates its documentation.
packages/razorback-plugin-dab/src/razorback_plugin_dab/cli.py	Changes CLI default --postgres-volume-mode to "fresh" and updates CLI help messaging.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.