retry arrivals / 20 ms bin
control ██████████ · · · ████████ · · · ██████ · · · ████
candidate ██████████ ▂ ▃ ▂ ▃ ▂ ▂ ▃ ▂ ▃ ▂ ▂ ▃ ▂ ▂ ▃ ▂ ▂ ▃
initial wave → post-failure arrivals →
Illustrative shape only, not a captured run. The first line keeps returning near shared boundaries. The second disperses later arrivals without pretending the initial synchronized trigger never happened.
Lab note: the observation
The policy executes in clients; the load shape appears at the dependency. A client log can show a random sleep value. It cannot show whether many independent clients still recombined into a burst at the last shared ingress point.
The artifact I want is therefore small: unsmoothed retry arrivals per time bin, split by cohort, with dependency saturation visible on the same clock. Original attempts stay out of the retry series, but the initial retry spike stays in. Removing it would make the chart prettier and the test less honest.
Controls worth freezing
Compare an unjittered capped-backoff cohort with one precisely described jittered candidate. Hold the synchronized trigger, attempt limit, backoff cap, timeout policy, destination, and offered load constant. Capture the retry ordinal and a test-run identifier with each destination-side arrival.
Do not call the candidate “Full Jitter” merely because someone used that label in configuration. The cap, sampling interval, timer granularity, and any state carried from the previous sleep have to come from the implementation itself. An arrival shape can support de-correlation; it cannot identify which randomization formula produced it.
Bin width is also a test parameter. It should be narrow enough to expose boundary-aligned peaks and wide enough not to turn timestamp noise into structure. Keep raw counts, then look at one neighboring width. Smoothing may help orientation, but it should not replace the bins that could falsify the claim.
What the source actually supports
Marc Brooker’s AWS Architecture Blog article on exponential backoff and jitter uses a synchronized optimistic-concurrency simulation against one remote database object. In that setup, capped exponential backoff reduces call frequency but leaves clusters separated by quiet periods. Adding jitter spreads calls toward an approximately constant rate beyond the initial spike.
The article’s 100-client simulation reports more than a 50 percent reduction in calls and better completion time for jittered backoff than for unjittered exponential backoff. Those values belong to that simulator, its network assumptions, and its OCC workload. They are not an acceptance threshold for an unrelated service.
It also compares Full, Equal, and Decorrelated Jitter. Full uses less work but slightly more time than Decorrelated in that comparison. That is not a universal ranking, and a named comparison does not identify the candidate implementation. This note therefore makes no judgment about which named algorithm the candidate implements or which algorithm is best.
Reading the shape without declaring victory
Repeated narrow peaks after the first wave are consistent with synchronization surviving. Before blaming the randomizer, segment by retry ordinal, client version, response class, and common boundaries such as timeout expiry or rate-limit reset. A shared scheduler or coarse timer can quantize otherwise independent choices.
A broad post-initial band is evidence in favor of de-correlation, but only that. Check total attempts, time to completion, tail completion, error mix, and dependency saturation. A smooth histogram can still accompany too much work or unacceptable delay.
A smaller secondary ridge is a reason to segment, not average. No visible difference may mean the policy was not active, the run lacked power, or unrelated network variance masked it. None of those observations establishes that jitter failed or succeeded.
Stop expansion when later peaks still coincide with saturation, throttling, or another dependency limit. The stop condition is editorial guidance for this experiment, not a threshold supplied by AWS.
Bring the destination-side histogram, attempt cohort, cap sequence, and saturation trace to the next discussion; let the arrival shape do the arguing.