policy
min(250 ms · 2ⁿ, 30 s)
total 1 min 31.75 s
Exponential
The default, and the one the index is laid out on. Each wait is twice the last until it reaches the cap, so a dependency that is down for a minute is asked nine times rather than nine hundred.
40 ms / px
| 0 | 250 ms | |
|---|---|---|
| 1 | 500 ms | |
| 2 | 1 s | |
| 3 | 2 s | |
| 4 | 4 s | |
| 5 | 8 s | |
| 6 | 16 s | |
| 7 | 30 s | |
| 8 | 30 s |
Doubling is the only schedule that is both cheap when the failure is transient and quiet when it is not. The first four attempts are over inside two seconds, which covers a timeout and a lock. The last four take a minute between them, which is long enough for a deploy to finish.
The successive-gap ratio here is exactly 2.000 until the cap, and the cap is the only place it is not. Hold this page against Fixed and the difference is a measurement, not an impression.
plates
