Performance Engineering & Gatling Glossary¶
Comprehensive reference of performance engineering, queuing theory, SRE reliability, and Gatling terminology.
A – C¶
- Amdahl's Law
- Formula modeling the maximum theoretical parallel speedup $S(N) = \frac{1}{\sigma + \frac{1-\sigma}{N}}$ constrained by the fraction $\sigma$ of work that must execute serially.
- Apdex (Application Performance Index)
- Industry-standard score from
0.00to1.00summarizing user satisfaction relative to latency target $T$: $\text{Apdex}_T = \frac{\text{Satisfied} + 0.5 \times \text{Tolerating}}{\text{Total}}$. - Assertion
- Simulation-wide pass/fail criterion evaluated at the end of a Gatling run (e.g.,
global().responseTime().percentile(95.0).lt(300)) that sets the CI exit status. - Backpressure
- Flow-control mechanism where a slow or saturated downstream receiver signals upstream producers to slow down rather than buffering unbounded queues in memory.
- Breakpoint Test (Capacity Test)
- Stepwise or ramped load test designed to locate the exact Knee of the Curve and Maximum Sustainable Throughput (MST) before SLOs breach.
- Bulkhead Pattern
- Isolating thread pools, connection pools, or semaphores per dependency or tenant so a single slow downstream service cannot sink the entire application.
- Cache Stampede (Thundering Herd)
- Failure mode where a hot cache entry expires under high concurrency, causing hundreds or thousands of simultaneous cache misses to overwhelm the backing database.
- Check
- Per-response validation or extraction rule in Gatling (e.g.,
status().is(200),jsonPath("$.id").saveAs("id")). Failed checks mark that request as KO. - Closed Workload Model
- Concurrency-driven model with a fixed cap of $N$ active users (
injectClosed); a new request only enters after a previous request completes. Appropriate for bounded worker pools, not open internet traffic. - Coordinated Omission
- Measurement flaw (identified by Gil Tene) where a load generator blocks or backs off during a server stall, omitting the requests that should have arrived during the freeze and drastically under-reporting tail latency (P99).
- Correlation
- Extracting dynamic data (tokens, IDs, CSRF values) from one response and injecting it into subsequent requests within the same virtual user's
Session.
E – L¶
- Error Budget
- Permissible unreliability over an SLO window (
100% - SLO). For a99.9%monthly SLO, the error budget is0.1%(43.8 minutesor1,000bad requests per million). - Feeder
- Data source (
csv,jsonFile, or custom JavaIterator<Map<String, Object>>) that injects distinct test records into virtual user sessions. - Four Golden Signals
- Google SRE's core service telemetry pillars: Latency (OK vs. KO separated), Traffic, Errors, and Saturation.
- Goodput
- Rate of successful, valid requests completing within SLO constraints per second (contrasted with raw throughput that may include fast
500/429errors). - HdrHistogram / DDSketch
- Fixed-memory logarithmic bucket data structures capable of recording millions of latency samples and computing accurate P99/P99.9 percentiles across merged nodes without averaging errors.
- Kingman's Formula (VUT Equation)
- Queuing approximation showing queue wait time grows as the product of Variability (arrival & service burstiness), Utilization ($\frac{\rho}{1-\rho}$), and Service Time ($S$).
- Knee of the Curve
- Inflection point on the throughput-vs-latency curve (typically ~70–85% utilization) where queuing delay shifts from flat/linear to steep exponential growth.
- Little's Law
- Fundamental queuing theorem $L = \lambda \times W$ (Concurrency = Arrival Rate $\times$ Residence Time), or $N = X \times (R + Z)$ when including think time $Z$.
- Load Shedding
- Deliberately rejecting excess requests fast (e.g.,
HTTP 429/503in< 2 ms) when saturated so admitted traffic continues meeting latency SLOs.
O – Z¶
- Open Workload Model
- Arrival-driven model (
injectOpen) where new virtual users arrive at rate $\lambda$ regardless of whether prior requests have finished—accurately modeling internet/API traffic. - Pacing (
pace) - Controlling iteration frequency by dynamically subtracting response time from a target cycle interval (
Wait = Pace - R), keeping RPS steady across minor latency shifts. - Percentile (
P50,P95,P99,P99.9) - Latency threshold below which $X\%$ of observations fall.
P99 = 800 msmeans 99% of requests finished in $\le 800\text{ ms}$ and the slowest 1% took $> 800\text{ ms}$. - RED Method
- Microservice monitoring framework tracking Rate (RPS), Errors (failed/sec), and Duration (latency distribution) on every endpoint.
- Session
- Gatling's immutable, per-virtual-user state map storing feeder variables, cookies, and correlated response values.
- SLI / SLO / SLA
- SLI = quantitative metric; SLO = internal reliability target; SLA = external legal/commercial contract carrying financial penalties.
- Soak Test (Endurance Test)
- Sustained load test over 8–24+ hours designed to expose memory leaks, connection leaks, file-descriptor exhaustion, and storage bloat.
- Spike Test
- Sudden extreme surge in arrival rate followed by return to baseline to verify load shedding, circuit breakers, auto-scaler lag, and clean recovery.
- Tail at Scale
- Microservice fan-out phenomenon where calling $k$ parallel backends amplifies tail probability according to $1 - (1 - p)^k$.
- Think Time (
pause) - Simulated human reading/interaction delay between sequential actions in a user journey.
- TTFB (Time to First Byte)
- Elapsed time between sending an HTTP request and receiving the first byte of the response header (isolates network RTT + server queuing/processing from body download).
- Universal Scalability Law (USL)
- Neil Gunther's law $C(N) = \frac{N}{1 + \alpha(N-1) + \beta N(N-1)}$ modeling both contention ($\alpha$) plateaus and coherency/crosstalk ($\beta$) retrograde throughput drops.
- USE Method
- Brendan Gregg's hardware/OS resource methodology inspecting Utilization, Saturation (queue length), and Errors for every CPU, disk, NIC, and pool.
- Volume Test
- Performance test executed against production-scale database table sizes (tens/hundreds of millions of rows) to uncover missing indexes and scan regressions.