Preparing the next view without exposing private workflow data.
Engineering task library · Version 5
Work that feels like work.
Design a system. Diagnose tail latency. Ship a migration. Recover a failed rollout. Pick a focused ticket or follow a project through its delivery phases.
Both adapters pass the happy-path tests, but the replacement sorts equal-name accounts differently and rounds stored revision values when decoding a database result. Parity dashboards currently compare only HTTP status codes.
All interactive tenants use the replacement, but a weekly synthetic export still imports the legacy repository directly. Deleting the old table based on request traffic would break that job and erase the only rollback copy.
The return endpoint accepts a caller-supplied status string. A replacement can be requested before inspection, and a stale browser tab can move a refunded return back to awaiting-inspection.
The browser retries an approved 4,500-minor-unit refund after a lost response. The current deduplication key is global, and changing the amount while reusing the key overwrites the queued request.
The application marks a return processing, then publishes a refund job. If it crashes between those steps, the return is stuck forever; publishing first can instead process a refund for a rolled-back decision.
A replacement reserves the last unit, then shipment creation fails. Retrying from the beginning makes another reservation, while immediately issuing a refund would ignore stock still held for the customer.
The synthetic refund provider begins timing out. The proposed circuit breaker maps every timeout to failed and releases queued retries when it closes, even though some timed-out refunds were accepted by the provider.
Stock reservations and refunds share a 12-slot provider pool. Twelve hanging stock requests prevent an otherwise healthy refund provider from receiving work, and the pending promise list grows without a limit.
A shipment-create response was lost, so the return entered compensation. A later status lookup confirms the replacement shipped; the queued compensation still plans to release stock and refund the customer.
Support can currently click Retry beside a return without seeing whether it will check a refund status, retry a reservation, or attempt compensation. The button also lets a stale page replay a command that already completed.
The happy-path demo succeeds, but no one has restarted the application between a provider accepting an operation and the database recording its result. A green HTTP test does not establish whether accepted refunds survive recovery correctly.
A ten-minute synthetic outage leaves 2,000 deferred returns. When the circuit closes, every due retry enters the refund pool together, crowding out new work and exhausting the shared database connection budget.
Build a calendar booking workflow with time-zone rules, conflict handling, idempotent provider calls, and safe reconnect behavior.
10 tickets · 3 phases
Take the brief into your own workflow.
For engineers
Practice scoped changes, keep a portable implementation and verification record, and learn to explain operational tradeoffs. Choose a ticket whose prerequisites you can provide.
For companies
Use realistic work to structure onboarding, internal practice, and conversations about engineering decisions. Each project names the delivery benefit. Agree scope and compensation before requesting company-specific work.
CSV contains one row per ticket. Map fields and issue types in your tracker; project grouping and dependency keys are descriptive. JSON preserves the complete project structure. These downloads do not synchronize with Jira.