Unverified Commit 004dc50d authored by Peter Goodspeed-Niklaus's avatar Peter Goodspeed-Niklaus Committed by GitHub
Browse files

Add metrics timing message passing from OverseerSubsystemContext to Overseer::route_message (#2201)

* add timing setup to OverseerSubsystemContext

* figure out how to initialize the rng

* attach a timer to a portion of the messages traveling to the Overseer

This timer only exists / logs a fraction of the time (configurable
by `MESSAGE_TIMER_METRIC_CAPTURE_RATE`). When it exists, it tracks
the span between the `OverSubsystemContext` receiving the message
and its receipt in `Overseer::run`.

* propagate message timing to the start of route_message

This should be more accurate; it ensures that the timer runs
at least as long as that function. As `route_message` is async,
it may not actually run for some time after it is called (or ever).

* fix failing test

* rand_chacha apparently implicitly has getrandom feature

* change rng initialization

The previous impl using `from_entropy` depends on the `getrandom`
crate, which uses the system entropy source, and which does not
work on `wasm32-unknown-unknown` because it wants to fall back to
a JS implementation which we can't assume exists.

This impl depends only on `rand::thread_rng`, which has no documentation
stating that it's similarly limited.

* remove randomness in favor of a simpler 1 of N procedure

This deserves a bit of explanation, as the motivating issue explicitly
requested randomness. In short, it's hard to get randomness to compile
for `wasm32-unknown-unknown` because that is explicitly intended to be
as deterministic as practical. Additionally, even though it would never
be used for consensus purposes, it still felt offputting to intentionally
introduce randomness into a node's operations. Except, it wasn't really
random, either: it was a deterministic PRNG varying only in its state,
and getting the state to work right for that target would have required
initializing from a constant.

Given that it was a deterministic sequence anyway, it seemed much simpler
and more explicit to simply select one of each N messages instead of
attempting any kind of realistic randomness.

* reinstate randomness for better statistical properties

This partially reverts commit 0ab8594c.

`oorandom` is much lighter than the previous `rand`-based implementation,
which makes this easier to work with.

This implementation gives each subsystem and each child RNG a distinct
increment, which should ensure they produce distinct streams of values.
parent 0253be89
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment