Implement Timeouts For Lua Scripts

[kevin-montrose]

For perf reasons this became a little more complicated than I would like, but I think it's justifiable.

Introduces LuaTimeoutManager which spins up a dedicated Thread. When SessionScriptCaches are populated (either for the first time, or after a previous Clear), they register themselves with LuaTimeoutManager. That registration is used to track when a Lua script is actually running, and to trigger best-effort timeouts.

The timeouts themselves use Lua's debug interface which provides a thread-safe way to clear/set hooks. This hook is set dynamically, because always having a hook registered causes a significant performance hit.

Performance issues are also why LuaTimeoutManager spawns a Thread. Using built-in timers or Tasks can hurt perf by putting more pressure on the shared thread pool - and most .NET timers also have some internal locking that hurts in practice.

To claw back some of the unavoidable overhead, I made three unrelated performance improvements:

• Assume that script hashes are lowercase (with a slower retry if they are not)
• Assume that ScriptHashKey.Equals will typically return true
• Add a fast path for MakeUpperCase to avoid walking byte-by-byte in the common case, as this is called unconditionally for scripting commands

Correctness

There are a couple concerns to handle in this design.

The first is to make sure we're not timing out spuriously - that's done by conceptually splitting the timeout into 10ths and waiting ~11 / 10ths before triggering a timeout. This deals with the case where registration happens in the middle of one of the 10ths.

The second is timeouts trigger after a Lua script has finished but before LuaTimeoutManager has been informed of that completion. If handled incorrectly, this could cause a subsequent script to be spuriously cancelled. This is handled by generating a "cookie" value when a script starts, and having it passed back on cancellation - if the cookies do not match expected values, the cancellation signal is ignored.

I added a stress test that runs 16 sessions in parallel with timeouts for an hour, and ran it for ~12 hours, to check the implementation. So it doesn't seem obviously broken, at least. This test is [Ignored] by default.

Performance

All runs are as of d148f51740d79548ad78dc19008fca0e58c0de43 for main and 28e1129eaba16cf60fdc101ed91d455e4a123df3 for luaTimeLimits.

ScriptOperations

Here timeouts are disabled for both branches, reporting just Native workloads.

There's a little bit of motion in either direction, but I'd say it's all within the margin of error.

main

Method	Params	Mean	Error	StdDev	Gen0	Gen1	Gen2	Allocated
ScriptLoad	Native,None	84.44 us	0.484 us	0.429 us	-	-	-	9600 B
ScriptExistsTrue	Native,None	19.27 us	0.313 us	0.384 us	-	-	-	-
ScriptExistsFalse	Native,None	18.66 us	0.140 us	0.131 us	-	-	-	-
Eval	Native,None	62.04 us	0.461 us	0.408 us	-	-	-	-
EvalSha	Native,None	27.86 us	0.184 us	0.163 us	-	-	-	-
SmallScript	Native,None	52.87 us	0.550 us	0.488 us	-	-	-	-
LargeScript	Native,None	4,147.26 us	25.975 us	23.026 us	-	-	-	-
ArrayReturn	Native,None	108.74 us	0.692 us	0.613 us	-	-	-	-

luaTimeLimits

Method	Params	Mean	Error	StdDev	Gen0	Gen1	Gen2	Allocated
ScriptLoad	Native,None	83.41 us	0.579 us	0.542 us	-	-	-	9600 B
ScriptExistsTrue	Native,None	19.67 us	0.177 us	0.157 us	-	-	-	-
ScriptExistsFalse	Native,None	17.95 us	0.195 us	0.182 us	-	-	-	-
Eval	Native,None	63.32 us	0.581 us	0.544 us	-	-	-	-
EvalSha	Native,None	26.76 us	0.121 us	0.107 us	-	-	-	-
SmallScript	Native,None	55.63 us	0.936 us	0.875 us	-	-	-	-
LargeScript	Native,None	4,079.82 us	42.402 us	39.663 us	-	-	-	8 B
ArrayReturn	Native,None	111.97 us	1.296 us	1.149 us	-	-	-	-

LuaTimeouts

A new benchmark focused on just timeouts, meant for checking improvements there. I'm benchmarking a 1 sec timeout (so checking for timeouts ever 100ms) and a 1 minute (so every 6 seconds there's a check), about the bounds of "reasonable" for timeouts.

These runs have 500K EVALSHAs (to tamp down variance). Naively there was about a 10% loss - but with some profiling and the aforementioned performance improvements I clawed that back. With timeouts enabled we're basically the same, with them disabled this is a ~2% improvement.

main

This was run by copying the benchmark on top of d148f51740d79548ad78dc19008fca0e58c0de43. Obviously the two with timeout cases are not included.

Method	Params	Mean	Error	StdDev
RunScript	Native,None	162.3 ms	1.53 ms	1.43 ms

luaTimeLimits

Method	Params	Mean	Error	StdDev
RunScript	Native,None,-	158.7 ms	1.09 ms	0.85 ms
RunScript	Native,None,00:00:01	159.6 ms	2.55 ms	2.26 ms
RunScript	Native,None,00:01:00	163.3 ms	1.78 ms	1.57 ms