THE FOLD / RESPAWN / SECOND WIND / THE JITTERED BACKOFF
THE JITTERED BACKOFF
a deterministic rule everyone shares is a coordination mechanism
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A hundred clients collide, all back off by the same doubling amount, and all return at the same instant to collide again. Exponential backoff without randomness does not spread load — it synchronises it.
LIT verified live. 100 clients, base 10 ms, cap 1 s, same seed, same collision rule. Pure exponential backoff drains 0 of 100 within a 200,000 ms window, burning 20,500 attempts, every one of which collides. Full jitter — wait a uniform random amount in
LIT verified live. 100 clients, base 10 ms, cap 1 s, same seed, same collision rule. Pure exponential backoff drains 0 of 100 within a 200,000 ms window, burning 20,500 attempts, every one of which collides. Full jitter — wait a uniform random amount in
[0, backoff) — drains 100 of 100 by 837 ms on 509 attempts, about 5.09 per client.2 HOW IT WAS WEAVED · AI + HUMAN
The comparison is Marc Brooker’s, in the AWS Architecture Blog piece on backoff and jitter; full jitter is the variant that wins there and here.
AVAN (AI) reports what happened rather than a ratio. The exponential arm never finished, so any speed-up figure would be a comparison against the length of the loop I chose, which is a property of my harness and not of the algorithm. 0 of 100 and 100 of 100 is the honest statement. The mechanism is not that jitter is faster; it is that identical clients running an identical deterministic rule remain identical forever, and randomness is the only thing that breaks the symmetry.
AVAN (AI) reports what happened rather than a ratio. The exponential arm never finished, so any speed-up figure would be a comparison against the length of the loop I chose, which is a property of my harness and not of the algorithm. 0 of 100 and 100 of 100 is the honest statement. The mechanism is not that jitter is faster; it is that identical clients running an identical deterministic rule remain identical forever, and randomness is the only thing that breaks the symmetry.
3 ONE DIMENSION
Retry instants. One arm is a comb; the other is a spread.
4 TWO DIMENSIONS · INTERACTIVE
Run both arms and watch who drains.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: a hundred clocks striking together.
AVAN’s addition (the inverse-companion): the forward reading is that jitter fixes retry storms. The inverse is that the storm was caused by the fix. Backoff was introduced to reduce contention, and doubling is the most reasonable-looking rule available — and because every client is running that same reasonable rule, it manufactures the very lockstep it was meant to prevent. Read backwards, randomness here is not a heuristic or a hedge; it is the only way a population of identical agents can ever stop agreeing, and a deterministic protocol shared by everyone is a coordination mechanism whether or not you wanted one.
LIT 100 clients, base 10 ms and cap 1 s on the same seed: pure exponential backoff drains 0 of 100 within a 200,000 ms window while burning 20,500 attempts, every one of which collides, whereas full jitter - a uniform random wait in [0, backoff) - drains 100 of 100 by 837 ms on 509 attempts, about 5.09 per client
FIG The comparison is Marc Brooker's, in the AWS Architecture Blog piece on backoff and jitter. AVAN reports what happened rather than a ratio: the exponential arm never finished, so any speed-up figure would be a comparison against the length of the loop I chose, which is a property of my harness and not of the algorithm. 0 of 100 against 100 of 100 is the honest statement. The mechanism is that identical clients running an identical deterministic rule remain identical forever, and randomness is the only thing that breaks the symmetry.
FIG The comparison is Marc Brooker's, in the AWS Architecture Blog piece on backoff and jitter. AVAN reports what happened rather than a ratio: the exponential arm never finished, so any speed-up figure would be a comparison against the length of the loop I chose, which is a property of my harness and not of the algorithm. 0 of 100 against 100 of 100 is the honest statement. The mechanism is that identical clients running an identical deterministic rule remain identical forever, and randomness is the only thing that breaks the symmetry.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of SECOND WIND · David Lee Wise (ROOT0), with AVAN