Design drills

The product cache outgrows its memory

A Redis cache with the default eviction policy meets a catalog three times bigger.

Sessions and cache share one Redis

Users are being logged out under load. Security wants sessions to survive restarts.

Moving a checkout script to Redis Cluster

A Lua script touches a user's cart and a global stock counter. Now it must run on a cluster.

A new service needs last week's orders

An analytics service joins an orders topic and finds it only sees new messages.

Orders must survive a broker failure

Replication factor 2, acks=1, and a requirement for zero acknowledged-message loss.

A slow fraud check joins the consumer

Each record now takes up to 800 ms, and every flagged order must be written exactly once.

Emails must not vanish when a worker crashes

One email worker with automatic acks is about to become four.

Upgrading a cluster that uses mirrored queues

Payments run on classic mirrored queues. The cluster is moving to RabbitMQ 4.

Ledger entries must stay in order per account

Eight consumers share one queue. Now each account's entries must apply in order.

Long video jobs join a queue built for images

A queue tuned for 5-second image jobs is about to carry 8-minute video jobs.

A FIFO queue meets Black Friday

Order updates use one message group. Traffic is about to grow ten times.

One bad record keeps retrying a whole batch

Lambda reads SQS in batches of ten. A rate-limited API now rejects single records.

A price cache with no eviction

A Hazelcast map caches prices with default settings. The catalog grows ten times.

Embedded members in an autoscaling app

Sessions live in embedded Hazelcast inside pods that scale every few minutes.

Flash-sale stock in a near-cached map

Stock counts are read through a near cache and updated with get then put.

Customers see yesterday's price

A product cache with a 24-hour TTL keeps showing old prices after an update.

The database spikes every minute

A hot homepage key expires every 60 seconds and thousands of requests rebuild it at once.

An old name comes back for a day

Writes update the database and delete the key, yet some users see stale profiles for hours.

Checkout fails when analytics is down

Checkout calls three services in a row before replying. A fourth team wants in.

Some orders never reach the warehouse

The order service commits to its database, then publishes to Kafka. Sometimes only one happens.

Orders span payment, restaurant and courier

A synchronous chain of calls needs manual fixes whenever a step fails halfway.