CodingLevel 6
Capacity and traffic
Easynot started
Mediumnot started
Hardnot started
Saturation & capacityWhy performance doesn't degrade linearly, it falls off a cliff.3 minRate limitingControl incoming traffic by issuing a fixed number of tokens over time.3 minMissing rate limitingWhen resource consumption bankrupts your company.3 minHot partition saturationWhen a billion-dollar distributed database gets taken down by one viral celebrity.3 minModelling the resources before the endpointsGet the nouns right and the endpoints mostly write themselves; get them wrong and you pay in round trips, in names nobody can say out loud, and in fields that mean two things.7 minDisk inode exhaustionWhen your server says "Disk Full" but you have 500GB of free space.3 minQueue backlogWhen work arrives faster than it's done, the line only gets longer.4 minLog volume managementWhy logging every HTTP request is the fastest way to bankrupt your startup.3 minThe storage stackEvery read and write travels down a ladder of layers, each one slower but more durable than the last. The whole craft of storage is keeping the data you reach for most in the fast layers near the top.4 minFalse sharing (cache contention)When adding more CPU cores makes your multithreaded code slower.3 minA write rejected when the quota runs outThe tenant's bucket is nearly full; the next write would spill over the line, so the store turns it away, and if you'd already started writing, you're left holding half an object.4 minEvent loop blockedWhy a single "while" loop can freeze Node.js for every user in the world.3 minGPU memory fragmentationWhy LLM inference engines used to waste 50% of their GPU memory, and how PagedAttention fixed it.3 minInference GPU autoscalerWhy you can't just spin up a new GPU server the second you get a traffic spike.3 minInference GPU OOM (out of memory)Why a single long document can crash an entire AI server.3 minLatency from a failing disk (triage, contain, root-cause)A dying disk rarely fails cleanly, it slows down first, dragging p99 latency up as it retries bad sectors. Triage the symptom, contain the blast radius, then prove the root cause.6 minShadow traffic overloadMirroring live requests to a new model is free safety, until the mirror eats the capacity the real traffic needs.3 minSnapshot bloat and disk pressureSnapshots are cheap to take but they keep old blocks alive, take enough and never prune, and they quietly eat the disk until writes start failing.4 minThrottling and 429s on object storageWhen you hammer one prefix too fast, the store says slow down, and the fix is to back off, not retry harder.3 minTraffic spikeWhen the flood arrives faster than you can add boats, you bail water at the door so the boat you have stays afloat.4 minWAL, tombstones, and read amplificationWrites are cheap appends; the cost is paid later, when one read has to dig through many layers, until compaction tidies up.4 minWhen the kernel kills your processMemory climbs toward a hard line. The moment it crosses, the kernel picks a victim and sends one quiet, fatal signal.3 min