| 1 |
--- |
| 2 |
title: Schedule randomization |
| 3 |
authors: ["@andriyDev"] |
| 4 |
pull_requests: [25094] |
| 5 |
--- |
| 6 |
|
| 7 |
Before a schedule runs (and therefore, your systems), it first computes the order that systems run |
| 8 |
in based on the ordering constraints (`.before()`, `.after()`, `.chain()`) of systems and system |
| 9 |
sets. However, in addition to this, the schedule must also resolve **conflicts** - if system A and |
| 10 |
system B both mutate component C, and there's no ordering between A and B, the schedule needs to |
| 11 |
pick one to run first. So far, the rule has been that this is non-deterministic. |
| 12 |
|
| 13 |
In practice though, schedules pick the order of these conflicting systems "deterministically, but |
| 14 |
arbitrarily". Put simply, your systems might accidentally be in the right order, but making an |
| 15 |
unrelated change to the graph might suddenly put it in the wrong order. This problem can be very |
| 16 |
difficult to detect. |
| 17 |
|
| 18 |
Introducing schedule randomization! This will randomize the order of systems while maintaining any |
| 19 |
explicit system ordering constraints. Once the `debug` feature is enabled, `ScheduleBuildSettings` |
| 20 |
will include a `shuffle_seed` field, that users can set to randomize their schedules. For example: |
| 21 |
|
| 22 |
```rust |
| 23 |
App::new() |
| 24 |
.add_plugins(DefaultPlugins) |
| 25 |
.edit_schedule(Update, |schedule| { |
| 26 |
// Make sure to add the `rand` crate with `cargo add rand`. |
| 27 |
let rng_seed: u64 = rand::random(); |
| 28 |
// Consider logging out the seed, so you can reproduce the error if you find a bug! |
| 29 |
info!("Randomizing Update schedule with seed={rng_seed}"); |
| 30 |
schedule.set_build_settings(ScheduleBuildSettings { |
| 31 |
shuffle_seed: Some(rng_seed), |
| 32 |
..Default::default() |
| 33 |
}); |
| 34 |
}) |
| 35 |
.run(); |
| 36 |
``` |
| 37 |
|
| 38 |
This can be used for "property testing", to verify that your systems satisfy some property despite |
| 39 |
different orderings of systems. |
| 40 |
|
| 41 |
There are some caveats however. Currently, when using `auto_insert_apply_deferred`, systems with |
| 42 |
commands are always placed before the earliest sync point they can. This means that although your |
| 43 |
systems may not have the correct ordering, they might "accidentally" have the correct ordering |
| 44 |
because of which sync point it uses. We hope to fix this in the future. |
| 45 |
|
| 46 |
In addition, the multi-threaded executor executes systems greedily: it looks for the first |
| 47 |
unexecuted system whose dependencies are finished and that has no other conflicting systems running. |
| 48 |
The result is that even if the shuffle results in the order `(A, B, C)`, `C` could run before `B` if |
| 49 |
`A` and `B` conflict. **This can be desirable to test**, but consider using the single-threaded |
| 50 |
executor to avoid this case. |