Small, self-contained demos of the systems ideas worth keeping sharp. Every one lets you drag the load up, change the strategy, and push it into the regime where it breaks — because the intuition lives at the cliff edge, not the happy path.
Fixed and sliding windows, token and leaky buckets — the same traffic hitting all four at once. Watch which absorb bursts, which smooth, which leak at the edges.
Add a node, watch only 1/N of the keys move. Virtual nodes fix the skew in front of you.
Naive retries amplify load during an outage. Add jitter, backoff, and a retry budget and watch it tame.
LRU vs LFU vs FIFO on different workloads, with a live hit rate. Spot the anomalies.
Round-robin vs least-connections vs power-of-two-choices. The last one is almost magic.
Little's Law made physical. Arrival past service rate, and the queue grows without bound.
Closed, open, half-open. Trip it, let it cool down, watch it probe for recovery.
Pick a bad partition key and watch one shard go red. Then shard the writes and cool it down.