Library
Concepts
46 terms the site keeps using, each in two or three plain sentences with a diagram and the places it shows up.
- At-least-once deliveryAt-least-once delivery means every message is delivered, and some may be delivered more than once.
- BackpressureBackpressure is a slow receiver telling a fast sender to slow down, instead of silently accepting more than it can handle.
- Blast radiusBlast radius is how much of a system one failure or one bad change can affect.
- Burn rateBurn rate is how fast a service is spending its error budget, compared with the pace that would use the budget up exactly at the end of the SLO period.
- Cache stampedeA cache stampede happens when a popular cached value expires and every request that arrives during the rebuild misses the cache.
- Cache-asideCache-aside is the most common way to use a cache: the application checks the cache first, and on a miss it reads the database and writes the value into the cache itself.
- CanaryA canary is a small part of the fleet that receives a new release first, with a small slice of real traffic.
- CardinalityCardinality is the number of distinct time series a metric produces: one for every unique combination of its label values.
- Circuit breakerA circuit breaker watches the calls to one dependency and stops making them once too many fail.
- Connection pool exhaustionConnection pool exhaustion is when every connection in an application's fixed pool is in use and new requests must wait for one.
- Consistent hashingConsistent hashing places both servers and keys on a ring, and each key belongs to the next server along.
- Consumer lagConsumer lag is how far a consumer is behind the newest message in a log or queue.
- Dead man's switchA dead man's switch is an alert that fires all the time on purpose, sent to a system outside the monitoring stack.
- Double-entry ledgerA double-entry ledger records every movement of money as one entry with two legs: a debit from one account and an equal credit to another.
- DownsamplingDownsampling replaces many raw data points with one summary per longer interval, such as one value every five minutes or every hour.
- Error budgetAn error budget is the amount of failure an SLO allows: with a 99.9% target, the budget is the other 0.1%.
- Exactly-once delivery (a myth)No system can promise that a message is delivered exactly once when networks and processes can fail, because the sender cannot tell a lost message from a lost acknowledgement.
- Exponential backoff with jitterExponential backoff means waiting longer after each failed attempt, typically doubling the wait.
- Feature flagA feature flag is a switch, read while the software runs, that decides whether a piece of code is used.
- Fencing tokenA fencing token is a number that goes up every time a lock or lease is granted.
- Head and tail samplingHead sampling decides whether to keep a trace as early as possible, when the request starts, so it is cheap but cannot know which requests will fail or be slow.
- Hot keyA hot key is a single key that receives a large share of all requests, such as one viral post.
- Hot partitionA hot partition is one partition or shard that gets far more data or traffic than the others.
- IdempotencyAn operation is idempotent if doing it twice has the same effect as doing it once.
- Kill switchA kill switch is a control that turns a feature or code path off everywhere, quickly, without deploying anything.
- Leader electionLeader election is how a group of nodes chooses the one node that will take writes and coordinate the rest.
- Load sheddingLoad shedding is refusing some requests on purpose so that the rest can succeed.
- Metastable failureA metastable failure is an outage that keeps itself going after the thing that triggered it has been fixed.
- MVCCMulti-version concurrency control lets readers and writers work at the same time without blocking each other.
- Outbox patternThe outbox pattern makes saving data and publishing an event about it behave as one step.
- p99 latencyp99 latency is the response time that 99 out of 100 requests beat.
- QuorumA quorum is the smallest number of replicas that must agree before a read or a write counts.
- Rate limitingRate limiting caps how many requests a client may make in a period of time.
- Read-your-writesRead-your-writes is the guarantee that after you change something, your own later reads show the change.
- ReconciliationReconciliation is comparing your own record of what happened with an independent one, such as your ledger against the bank's statement.
- Replication lagReplication lag is the delay between a change being confirmed on the primary database and that change appearing on a replica.
- Retry stormA retry storm is a flood of repeated requests set off by failures, which then causes more failures.
- SagaA saga is a business transaction that spans several services, carried out as a sequence of local transactions.
- Shuffle shardingShuffle sharding gives each customer its own small, randomly chosen set of workers out of a larger pool.
- SLOA service level objective is a target for how well a service should behave, such as 99.9% of requests succeeding over 30 days.
- Split brainSplit brain is when two parts of a system each believe they are in charge and both accept writes.
- Thundering herdA thundering herd is a large number of clients or requests all doing the same thing at the same moment.
- Token bucketA token bucket is a rate limiter that gives each client a bucket of tokens, refilled at a steady rate up to a fixed size.
- TTL jitterTTL jitter is a small random amount added to each cache entry's time-to-live.
- Virtual nodesVirtual nodes place each server at many points on a hash ring instead of one.
- Write-throughWith a write-through cache, every write goes to the cache and to the database before the caller is told it succeeded.