Michael Sproul 58012f85e1 Shutdown gracefully on panic (#2596) ... ## Proposed Changes * Modify the `TaskExecutor` so that it spawns a "monitor" future for each future spawned by `spawn` or `spawn_blocking`. This monitor future joins the handle of the child future and shuts down the executor if it detects a panic. * Enable backtraces by default by setting the environment variable `RUST_BACKTRACE`. * Spawn the `ProductionBeaconNode` on the `TaskExecutor` so that if a panic occurs during start-up it will take down the whole process. Previously we were using a raw Tokio `spawn`, but I can't see any reason not to use the executor (perhaps someone else can). ## Additional Info I considered using [`std::panic::set_hook`](https://doc.rust-lang.org/std/panic/fn.set_hook.html) to instantiate a custom panic handler, however this doesn't allow us to send a shutdown signal because `Fn` functions can't move variables (i.e. the shutdown sender) out of their environment. This also prevents it from receiving a `Logger`. Hence I decided to leave the panic handler untouched, but with backtraces turned on by default. I did a run through the code base with all the raw Tokio spawn functions disallowed by Clippy, and found only two instances where we bypass the `TaskExecutor`: the HTTP API and `InitializedValidators` in the VC. In both places we use `spawn_blocking` and handle the return value, so I figured that was OK for now. In terms of performance I think the overhead should be minimal. The monitor tasks will just get parked by the executor until their child resolves. I've checked that this covers Discv5, as the `TaskExecutor` gets injected into Discv5 here: f9bba92db3/beacon_node/src/lib.rs (L125-L126)		2021-09-15 00:01:18 +00:00
..
environment	Update sloggers to v2.0.2 (#2588)	2021-09-14 06:48:26 +00:00
src	Shutdown gracefully on panic (#2596)	2021-09-15 00:01:18 +00:00
tests	Shutdown after sync (#2519)	2021-08-30 13:46:13 +00:00
Cargo.toml	Update sloggers to v2.0.2 (#2588)	2021-09-14 06:48:26 +00:00