Home

Glossary

Every term the playground uses, in plain English.

Alert
An automatic notification that a metric crossed a threshold for a while.
API gateway
The front door for API traffic: authentication, rate limits and routing.
Backlog / consumer lag
How many messages are waiting in a queue, or how old the oldest one is.
Cache
Fast memory that stores recent answers so the slow source (database) is asked less often.
Cache stampede
Many keys expire at once, so a flood of misses hits the database at the same moment.
Canary deploy
Releasing a new version to a small slice of servers first and watching it before continuing.
Capacity
The most throughput a component can sustain: roughly concurrency × 1000 ÷ time per request (ms).
Time spent waiting on dependencies counts too, because the slot is held while waiting.
Cascading failure
One slow or broken component makes its callers slow or broken, and so on up the chain.
CDN
Content Delivery Network: caches responses on servers close to users.
Circuit breaker
Stops calling a dependency that keeps failing, and fails instantly instead, then tries again later.
Cold cache
A cache with nothing in it, usually after a restart. Every read misses.
Concurrency
How many requests a component can work on at the same time: threads, workers, or connections.
Connection pool
The limited set of connections a database accepts at once. Queries wait when it is full.
Dead-letter queue
Where messages go after failing several times, so they do not block everything else.
Eventual consistency
Copies of data agree eventually, not instantly. You might briefly read stale data.
Exponential backoff
Waiting longer before each retry: 100ms, 200ms, 400ms…
Headroom
Spare capacity kept free for spikes and failures.
Health check
A periodic 'are you alive?' probe. Between checks, a dead server can still receive traffic.
Hit ratio
The fraction of reads a cache answers itself. The rest (misses) go to the next layer.
Hot key / hot shard
One key (or shard) that gets far more traffic than the others.
Jitter
Randomness added to wait times so that many clients do not all retry at the same instant.
Latency
How long one request takes, from sending it to getting the answer.
Usually measured in milliseconds (ms). Users notice anything above ~100ms and get annoyed above ~1s.
Load balancer
Spreads requests across replicas and stops sending to ones that fail health checks.
Load shedding
Deliberately rejecting some work when overloaded so the remaining work succeeds.
Logs
Timestamped messages written by components describing what they did or what went wrong.
Message queue
A buffer between producers and consumers. Work is accepted instantly and processed later.
Turns spikes into a steady stream. The price is that work finishes later than it was accepted.
Metrics
Numbers over time: request rate, errors, latency, utilisation.
Network partition
Two components can no longer reach each other, although both are running.
Noisy neighbour
Another program on the same physical machine using resources you expected to have.
p50 / p99
Percentiles of latency. p99 = 300ms means 99 of every 100 requests finish in under 300ms.
Averages hide pain. The slowest 1% are often your most active users, because they make the most requests.
Packet loss
Some network messages never arrive. The sender only finds out by timing out.
Queue (wait queue)
Where requests wait when every slot is busy. A growing queue means demand exceeds capacity.
Rate limiting
Rejecting requests above a set rate, quickly and cheaply, to protect the rest of the system.
Read replica
A read-only copy of the database. Takes read load off the primary, but may lag behind it.
Replica
An identical copy of a component. More replicas give more capacity and survive a crash.
Replication lag
How far a replica is behind the primary. Reads from it may return old data.
Retry
Trying a failed call again. Great for blips, dangerous when the dependency is overloaded.
Retry storm
Retries add load, which causes more failures, which cause more retries.
Saturation
A component running at (or near) 100% of its capacity. Requests start to wait.
Sharding
Splitting data across several databases by key, so each handles only part of the load.
Single point of failure
A component with no backup. If it fails, everything that depends on it fails.
SLO
Service Level Objective: the target you promise, e.g. p99 under 300ms and fewer than 1% errors.
Throughput (RPS)
How many requests are handled per second.
Thundering herd
Many clients or jobs doing the same thing at exactly the same moment.
Timeout
How long a caller waits for an answer before giving up.
Trace
The path of one request through every component it touched, with timings for each hop.
Utilisation
What fraction of a component's capacity is in use right now.
Latency climbs sharply as utilisation approaches 100%. Most teams aim for 50–70% at peak.