Every term this site uses, in one place. Each card opens the same drill-down as the underlined phrases in the text: one line, why it matters here, what it is, what it is not, and how it shows up in the insurance example.
Three sentences hold the words apart. A retry is not a recovery. A timeout tells you nothing about whether the write happened. The model recommends; a person decides.
Acceptable outcomeAn end state of a run that the insurer can live with, even if the happy path did not happen.AcknowledgementThe short message a worker sends to the queue when it has finished with a message, so that the queue deletes it.AlertA notification that asks a person to act now, which on this site fires when the error budget is burning too fast, not on every error.Amazon SQSA managed AWS queue in which a received message is hidden for a visibility timeout and comes back if the worker does not delete it.Apache KafkaA distributed log: producers append messages to partitions, and each consumer keeps its own position and moves it forward as it works.Approval timeoutThe longest a review task may wait for a decision before the system acts on the delay, by reminding or reassigning, never by deciding.At-least-once deliveryA delivery guarantee: every message is processed one or more times. Nothing is lost; a message may arrive again.At-most-once deliveryA delivery guarantee: every message is processed zero or one times. Nothing is duplicated; a message may be lost.attempt_numberA counter for the tries of one step inside one run: 1 for the first try, 2 for the first retry, and so on.AvailabilityThe system answers when asked: within the deadline it produces an end state instead of hanging or refusing.AWS Step FunctionsAn AWS service that runs state machines written in JSON, with Retry and Catch fields on each step and a callback token for outside answers.BackoffThe pause before a retry, which grows after every failure so that a struggling dependency gets time to recover.BackpressureLetting the slowest part of a pipeline slow down the parts that feed it, instead of piling work up in hidden buffers.Blast radiusHow much one failure can affect: one attempt, one step, one run, every run on a dependency, or every submission.Blue-green deploymentKeep two complete environments, deploy the new release to the idle one, and move the work across by switching the router; switch back to roll back.BudgetThe most a run, or all runs in a period, may spend, checked before each spending call so the ceiling is a limit and not a report.BulkheadCapping how much of a shared resource (workers, connections, threads) one dependency may occupy, so its failure cannot exhaust the rest.Burn rateThe bad share in a window divided by the bad share the objective allows: 1 spends the budget exactly over the period, 2 in half of it.CanaryA partial, time-limited release that sends a small share of the work to the new version and compares its results with the rest before going further.Capacity planningMaking sure there is enough capacity, and enough to spare, to serve the demand you expect at the reliability you promised.Chaos engineeringThe practice of running controlled fault experiments on a system to build confidence that it withstands turbulent conditions.CheckpointThe saved state of a run after one of its steps completed, written outside the process so a crash does not lose it.Circuit breakerA component in front of one dependency that stops sending calls to it after repeated failures, refuses calls at once for a while, and then lets a probe through.CompensationWhen a later step fails for good, an action that undoes what an earlier step meant, without erasing that it happened.Concurrency limitA cap on how many calls to one dependency may be in flight at once; the next caller waits or is refused until one finishes.CorrectnessThe answer is right and complete by the rules that apply, not only present.Cost ceilingThe most a single run may spend; a call that would cross it is not made, and the run ends with an honest escalation.Dead-letter queueA separate queue that holds messages a consumer could not process after a set number of tries, so they stop circulating and a person can look at them.DeadlineA point in time by which the whole run must be finished. Every call inside the run gets at most the time that is left.DeduplicationNoticing that a request, message or write has been seen before, and not doing its work a second time.Degraded resultA result given because a dependency was unavailable, smaller or from a less trusted route, and labelled as such on the stored record.DependencyAnything outside a step that the step needs in order to do its work.Distributed lockA record in a shared store that says which worker currently owns a resource, such as a submission.Duplicate writeWhat happens when a repeated write is not recognized as a repeat: two rows, two tasks or two emails for one intention.DurabilityOnce the system has accepted a request or written a result, it keeps it, through crashes and restarts.Durable executionA way of running long code so that the outcome of each completed step is recorded durably and a crash only repeats the step in progress.Error budgetOne minus the objective, times the number of events: how many bad events the period may contain before the objective is missed.Escalation pathThe ordered list of roles, each with its own approval timeout, that a review task moves through when nobody decides it in time.Exactly-once deliveryThe wish that every message is processed one time, never zero and never two. For a consumer whose work changes something outside the queue, a queue alone cannot provide it.Exactly-once effectHowever many times a request or message arrives, its effect is applied once.Fail fastRefuse immediately instead of waiting for a timeout to tell you what you could already know.Failure modeOne specific way in which a dependency fails, with its own signature and its own right response.FailureThe system as a whole did not deliver an acceptable outcome: no recommendation, two, or a broken one.Fault injectionCausing a chosen fault at a chosen place, on purpose, to see whether the recovery code works.Fault planA list of rules, written as data, that says which calls of which dependency fail and in what way.FaultA cause: something that went wrong inside a component, such as an overloaded server or a dropped packet.Feature flagA setting read at run time that switches a behaviour on or off, for everyone or for a share, without a new deployment.Fencing tokenA number that increases with every lease granted, sent with every write; the store refuses any write carrying a number lower than one it has already seen.Game dayA planned session in which a team causes failures on purpose and watches what the system and the people do.Graceful degradationGiving less than the full answer, chosen on purpose and labelled as such, instead of no answer or an answer that pretends to be complete.Gray failureA dependency that is not working properly while the checks that watch it report it as healthy.Half-openAfter the open time, the breaker lets a small number of probe calls through; success closes it, failure opens it again.HeadroomThe capacity left over at the expected peak, which decides how fast a backlog drains after something goes wrong.Hedged requestSending a second copy of a request when the first has taken longer than usual, using whichever answer comes first and cancelling the other.Idempotency keyA name the caller attaches to a write and repeats on every attempt, so the store can recognize a repeat and return the stored result.Idempotency-Key headerA proposed HTTP request header in which the client sends a key so the server can recognize a repeated request.IdempotentAn operation is idempotent when doing it twice leaves the same result as doing it once.In-flight runA run that is still between its first step and a terminal state when something changes around it, such as a deployment.InboxA record, kept in the consumer's own store, of the message ids it has accepted, so that a redelivered message does no work twice.IncidentAn unplanned event that harms or threatens the outcome for users badly enough that people must coordinate a response.InvariantA property that must hold after every run, checked only on what is left, independent of the fault that was injected.JitterA random part of the backoff wait, so that clients who failed at the same moment do not retry at the same moment.Kill switchA long-lived flag kept for emergencies, changed without a release, that stops a behaviour immediately.LangGraphA library for building agents as graphs of nodes, with checkpointers that save a thread's state so a run can continue after a failure.LeaseA lock granted for a limited time, which the holder must renew; if it does not, someone else may take it.Load sheddingWhen a system nears overload, rejecting some work early, in a chosen order, so the work it accepts is served well.Lost acknowledgementThe other side committed the change, but its answer was lost on the way back, so the caller sees an error.Malformed outputA model answer that arrived successfully and fails validation: not valid JSON, a value that is not allowed, a missing field, a citation to a rule that does not exist.MitigationAn action that reduces the harm of an incident now, whether or not it fixes the cause.Model fallbackSending the same step to a different model (another size, version or provider) when the primary route is down, so the run can continue.Model routeA provider, endpoint or region, model id and settings (prompt version, timeout, breaker) that together make one way of calling a model.OutboxA table in the service's own database where an outgoing event is inserted in the same transaction as the business change, for a separate relay to publish later.OverrideWhen the underwriter decides differently from what the agent recommended, recorded with who decided, what and why.Partial failureOne part of a distributed system finished its work while another part was cut off and cannot tell what happened.PercentileThe p-th percentile is the smallest value with at least p percent of the measurements at or below it; p50 is the median.Permanent errorAn error that will never pass by itself: the same request meets the same refusal every time.Poison messageA message that can never be processed, because the fault is in the message itself, so every delivery fails the same way.Policy fixtureThe illustrative underwriting guide that the agent retrieves and the rules check against, with versions.PostmortemA written account of an incident: impact, actions taken, causes and follow-up actions, focused on the system rather than on people.PromotionTurning what an incident taught into permanent tests: a fault scenario for the workflow and, when an answer was wrong, a golden case for evaluation.Provider failoverSending the same request, to the same model, through another region or endpoint when the first one is down.Quality impactThe difference in content between an answer from the fallback route and the answer the primary route would have given.QueueA place where a producer leaves work items (messages) and workers take them when they have capacity, so that a crashed worker's work is not lost.Queueing delayThe time a piece of work waits in a queue before a worker, a provider or a person starts on it.RabbitMQA message broker in which consumers acknowledge deliveries; unacknowledged deliveries are requeued and redelivered when a channel closes.Race conditionTwo activities touch the same thing, and the result depends on which one gets there first.Rate limit errorA refusal that says the caller sent too many requests in a window and should come back later, often with how long to wait.Rate limitA cap on how many calls may be made per second (or minute, or day), enforced by the provider, by the client, or both.RecommendationThe agent's output: approve, decline, request_info or escalate, with reasons, cited rules, anything missing, and requires_human: true.ReliabilityA system keeps delivering acceptable outcomes despite failures, retries, load and changing dependencies.ResilienceThe system keeps its outcomes acceptable while one of its dependencies is failing.Resilience4jA Java fault-tolerance library that wraps a call in decorators: retry, circuit breaker, rate limiter, bulkhead, time limiter and cache.ResumeLetting a new process pick up a run where the last one stopped, by loading its checkpoint.Retry budgetA limit on how many retries a process or system may send, so that retries can never grow into a multiple of normal traffic.Retry stormWhat happens when many clients retry a shared dependency at once: the retries add load exactly when the dependency can least carry it.Retry-AfterAn HTTP response header, defined in RFC 9110, that tells the client how long to wait before a follow-up request: an HTTP date or a number of seconds.RetryTrying the same step again, in the same run and the same process, after an error that may pass with time.Review taskA task in the underwriter cockpit that links one stored recommendation to the underwriter who must decide on it.RollbackGoing back to the previous release record after a change turned out worse, ideally automatically when an alarm fires.Root causeA cause that, if it had been absent, would have prevented the incident; complex incidents usually have several.RunOne execution of the eight-step workflow for one submission, named by its run_id.run_idThe identifier of one run, such as run-7f3a, under which its checkpoint, its trace and its idempotency keys are kept.RunbookA written procedure for a known failure: how to confirm it, what to do in which order, what never to do, and how to tell it worked.SagaA long piece of work split into steps that each commit on their own, where every step that can be undone has a compensation to run if a later step fails for good.Side effectA change outside the run that stays after the run ends: a row, a task, an email.SLAAn agreement with users or customers that names what happens when an objective is missed.SLIA measurement of one aspect of a service, best written as good events divided by all events, so it runs from 0 to 100 percent.SLOThe share of good events a service must reach over a stated period, agreed in advance.Steady stateA measurement of what the system delivers when it works normally, used as the baseline of a fault experiment.SubmissionOne structured record made from a broker's quote email; every later step reads it in place of the email.Synthetic scenarioA made-up but realistic input, with a fault plan and an expected end state, run through the real workflow.Tail latencyThe latency of the slowest share of requests or runs, read from high percentiles such as p95, p99 and p99.9.TemporalA platform that runs workflows as code, records an event history, and rebuilds a run's state by replay after a failure.Terminal stateA state a run ends in on purpose, which a person or a dashboard can read without guessing.Time to recoverHow long it takes from the moment outcomes start to suffer until the service delivers acceptable outcomes at its normal rate again.TimeoutThe longest we are willing to wait for one attempt of one call. After that we stop listening and treat the attempt as failed.Token bucketA bucket that gains tokens at a fixed rate up to a maximum; each call takes one, and a call with no token waits or is refused.Transient errorAn error that passes with time: trying again later may succeed.Unacceptable outcomeAn end state the business cannot live with: a duplicate write, a lost run, a recommendation without its evidence, a decision without a person.Underwriter cockpitThe application in which underwriters see review tasks, read the agent's recommendation and record their decision.UnderwriterThe person at the insurer who decides whether to accept a risk. The agent only prepares a recommendation for them.Underwriting agentThe program in the example that reads a submission, gathers evidence, has a model draft a recommendation, stores it and opens a review task.Version skewA run, or two cooperating systems, using parts from two releases at the same time, for example a run started on v14 and finished on v15.Visibility timeoutThe time a queue hides a received message from other workers. If no acknowledgement arrives in that time, the message becomes visible again.Workflow stepOne unit of the workflow, with a fixed id, an input from the steps before it, at most one kind of side effect and a rule for failure.
Each term is explained in full on one page and used everywhere else.