Consistency, replication & consensus

Slice the network between two nodes and choose, per request, between rejecting it (CP) and answering with possibly stale data (AP), then remove the partition and see the latency-vs-consistency trade that remains.

Watch a reply appear before its parent comment under eventual consistency, then switch to causal and see ordering restored.

Dial up replica lag, post a comment, refresh, and see it missing until sticky routing or version tokens fix it.

Move the point where the API returns 200 along the pipeline and crash things after it to show what the user was actually promised.

Switch topologies and turn R/W/N quorum knobs to see when reads are guaranteed to overlap the latest write.

Kill the leader, promote a lagging replica, and watch acknowledged writes disappear (RPO) while the outage clock runs (RTO).

A partitioned old leader keeps accepting writes, and you see how fencing or quorum-based promotion prevents two leaders.

Climb the three tiers and watch each one's outage behavior, and the new costs it adds: conflicts, global idempotency, and complicated consistency semantics.

Make concurrent edits in two regions and compare last-writer-wins, merge functions, and CRDT counters.

Step through votes and log replication one message at a time, forcing split votes, partitions, and stale leaders.

Stretch a GC pause past a lock's TTL to get two writers, then add fencing tokens and watch the stale write rejected.