[amalsaidov]

Motivation

Backoff.decorrelatedJittered is meant to produce a decorrelated jitter stream where each backoff is drawn from a range up to 3x the previously selected value, capped at maxMs (see the method javadoc and http://www.awsarchitectureblog.com/2015/03/backoff.html).

The conversion of the previous backoff into nanoseconds has its TimeUnit arguments swapped:

long lastNanos = TimeUnit.MILLISECONDS.convert(lastMs, TimeUnit.NANOSECONDS);

TimeUnit.MILLISECONDS.convert(d, NANOSECONDS) interprets d as nanoseconds and returns milliseconds, so it computes lastMs / 1_000_000. Since lastMs is already in milliseconds, lastNanos collapses to 0 for any sub-second backoff.

With lastNanos stuck at 0, randRange = abs(lastNanos * 3 - startNanos) is always startNanos, so every emitted backoff stays in [startMs, 2 * startMs) no matter how many steps elapse. The stream never grows toward maxMs and the decorrelation the strategy is supposed to provide is lost. The other jitter strategies in the same class (exponentialJittered, equalJittered) convert ms to ns correctly with TimeUnit.NANOSECONDS.convert(..., MILLISECONDS).

Changes

• Backoff.decorrelatedJittered: convert the previous backoff from milliseconds to nanoseconds using TimeUnit.NANOSECONDS.convert(lastMs, TimeUnit.MILLISECONDS), matching how startMs/maxMs are converted a few lines above.
• Added TestBackoff.testDecorrelatedJitteredGrowsBeyondStart, which fails before the fix (every backoff capped below 2 * startMs, observed max 19 for startMs=10, maxMs=10000) and passes after.

Verified by building and testing only the affected module:

mvn -pl bookkeeper-common -am test -Dtest=TestBackoff

AI-assisted; directed, reviewed and tested by the author.