Build a Resettable Practice Sandbox
Practice on production teaches fear; practice on a fake that drifts teaches wrong habits. A seeded container with a one-command reset makes every mistake cheap and every session reproducible.
Procedure
- Define the environment in
docker-compose.yml: app, database, and any queue, all pinned to explicit image digests. - Seed deterministic data in
seed.sqlwith fixed ids and timestamps, so learners can share expected answers. - Provide the reset as one target:
make reset, which runsdocker compose down -v,up -d, then loadsseed.sql. - Print a readiness check the learner can trust:
psql -tAc 'select count(*) from users'should print100. - Never mount real credentials or a production network; the compose file must declare its own network.
- Add a "break it" task: drop a table, run
make reset, confirm the learner is back to green. - Document the reset command at the top of the exercise
README.md, not buried in theMakefile. - Pin the seed date so time-dependent queries return the same rows tomorrow.
- Cap container memory and CPU so a runaway practice loop cannot freeze the laptop.
- Ship a smoke test
make checkthat fails if the seed count is wrong, so a broken sandbox is caught early. - Add a
.dockerignoreso builds are fast and no local secrets are copied in. - Name the database and volume explicitly so
down -vtargets the right one. - Print the reset instructions on a failed smoke test, so the fix is one command away.
- Document the exact reset command in one line at the top of the README.
- Keep the seed small enough to load in seconds.
- Log the container start time so a stuck boot is obvious.
Pitfalls
- Seeding with
now()or random ids, so two learners get different answers. - A reset that drops the volume but not the schema, leaving stale state behind.
- Pointing the sandbox at a shared staging database "just for convenience".
- Requiring a dozen manual steps to reset, so learners avoid experimenting.
- Leaving a real API key in the compose file for realism.
- Pin tags instead of digests, so a rebuild silently pulls a new image.
- Binding to a port already in use on the laptop, so startup fails confusingly.
- A seed script that depends on the wall clock.
- Assuming
dockeris installed without checking during bootstrap. - Reset buried under three make targets so nobody finds it.
- A seed so large that resetting interrupts practice.
- A silent hang on startup that looks like a broken exercise.
Verification
make reset && psql -tAc 'select count(*) from users'
# passes when the count is exactly 100 after reset, and drops again after a deliberate break
Report to the user: the reset command, the readiness check, and the pinned image digests.