Timeouts, backoff, and parallelism

Q: How do I implement exponential backoff in C#?

Use Pineapple’s ExponentialWait: var wait = new ExponentialWait(initialMs: 100, growth: 2, maxDelay: 5000); and call await wait.WaitAsync() between attempts. The nth wait is initialMs × (n + 1)^growth, capped at maxDelay.

Timeouts

AccurateTimeout and SimpleTimeout (in Pineapple.Threading) track a deadline and throw TimeoutException from EnsureNotTimedOut() once it passes. Both implement ITimeout.

await using var timeout = new AccurateTimeout(timeoutInMs: 30_000);

while (!await IsReadyAsync())
{
    timeout.EnsureNotTimedOut();                    // throws TimeoutException after 30 s
    await Task.Delay(500);
}
Console.WriteLine($"Ready after {timeout.ElapsedInMilliseconds} ms");

Type	Clock	Notes
`AccurateTimeout`	`Stopwatch` (monotonic, high resolution)	Preferred
`SimpleTimeout`	`DateTime.Now`, whole seconds	Lower overhead, coarse

ITimeout exposes HasExpired, ElapsedInMilliseconds, RemainingInMilliseconds, and EnsureNotTimedOut().

Exponential backoff

ExponentialWait delays for initialMs × (attempt + 1)^growth milliseconds, optionally capped at maxDelay.

var wait = new ExponentialWait(initialMs: 100, growth: 2, maxDelay: 5_000);
// waits: 400, 900, 1600, 2500, 3600, 4900, 5000, 5000 ... ms

for (var attempt = 0; attempt < 8; attempt++)
{
    if (await TryConnectAsync()) break;
    await wait.WaitAsync();
}

Parallelism

Parallelism computes a CPU-aware MaxDegreeOfParallelism — by default two-thirds of Environment.ProcessorCount (rounded down, minimum 1), and never more than the value you request.

IParallelism parallelism = new Parallelism(requestedDegreeOfParallelism: 8);
Parallel.ForEach(files, parallelism.Options, Process);

var heavy = new Parallelism(factor: 0.5m);   // use half the CPUs

Questions about timeouts and backoff

How do I implement exponential backoff in C#?

Use Pineapple’s ExponentialWait: var wait = new ExponentialWait(initialMs: 100, growth: 2, maxDelay: 5000); and call await wait.WaitAsync() between attempts. The nth wait is initialMs × (n + 1)^growth, capped at maxDelay.

How many threads should Parallel.ForEach use?

A common choice is a fraction of the logical processors. Pineapple’s Parallelism defaults to two-thirds of Environment.ProcessorCount (rounded down, minimum 1) and returns ready-made ParallelOptions.