Add parallel setup driver for GCP resources - #67

Adds scripts/setup_resources.sh, which runs the three setup scripts concurrently and reports a status table. Wall clock becomes Composer's ~25 min instead of the sum of all three.

Mirrors destroy_resources.sh in flags and structure, so the create/destroy pair stays symmetric.

Usage

./scripts/setup_resources.sh                        # all three in parallel
./scripts/setup_resources.sh --skip-composer        # skip the slow one
./scripts/setup_resources.sh --feature-store-args="fraud fraud_user"
./scripts/setup_resources.sh --serial               # one at a time, for debugging
./scripts/setup_resources.sh --dry-run

Output

=== Setup Summary ===
RESOURCE             STATUS     DURATION   LOG
--------             ------     --------   ---
composer             SUCCESS    24m18s     .setup-logs/20260807-135736/composer.log
feature-store        FAILED(7)  0m02s      .setup-logs/20260807-135736/feature-store.log
artifact-registry    SUCCESS    0m41s      .setup-logs/20260807-135736/artifact-registry.log

Exit code is 0 only if every target that actually ran succeeded.

Design notes

• Per-script log files. Three interleaved streams of gcloud output on one terminal is unreadable. Failed logs are tailed automatically in the summary.
• Deliberately not set -e. A failing child script has to be recorded and reported next to the others, not abort the driver and hide the rest.
• Pre-flight checks verify gcloud auth and that feature_store is importable. Without them, a missing venv fails feature-store one second in while Composer keeps building for 25 minutes.
• Skipped targets appear in the summary. A table that silently omits them reads as "all three succeeded" when only one ran.
• Accurate per-task durations — each task's finish time is stamped when detected (2s poll), rather than reading the clock after the whole loop exits, which would give every task the slowest one's duration.
• bash 3.2 compatible (macOS /bin/bash): indexed arrays only, no wait -n, no associative arrays.
• --timeout kills the process group rather than the parent, since gcloud spawns children that would otherwise outlive it. Off by default — interrupting a half-created Composer environment is worse than waiting.

Testing

Exercised against stub scripts covering: all-success, partial failure with exit-code propagation, skip reporting, serial mode, timeout watchdog, dry-run, unknown-argument rejection, all-skipped, and pre-flight pass. Not yet run against live GCP.

🤖 Generated with Claude Code