And maybe the webhook authorizer cache.
This cache has two primary advantages over the LRU cache used currently:
- Cache hits don't acquire an exclusive lock.
- More importantly, performance doesn't fallover when the access pattern
scans a key space larger than an arbitrary size (e.g. the LRU
capacity).
The downside of using an expiring cache here is that it doesn't have a
maximum size so it's suspectible to DoS when the input is user
controlled. This is not the case for successful authentications, and
successful authentications have a natural expiry so it might be a good
fit here.
It has some a few differences compared to:
3d7318f29d/staging/src/k8s.io/client-go/tools/cache/expiration_cache.go
- Expiration is not entirely lazy so keys that are never accessed again
are still released from the cache.
- It does not acquire an exclusive lock on cache hits.
- It supports per entry ttls specified on Set.
The expiring cache (without striping) does somewhere in between the
simple cache and striped cache in the very contrived contention test
where every iteration acquires a write lock:
```
$ benchstat simple.log expiring.log
name old time/op new time/op delta
Cache-12 2.74µs ± 2% 2.02µs ± 3% -26.37% (p=0.000 n=9+9)
name old alloc/op new alloc/op delta
Cache-12 182B ± 0% 107B ± 4% -41.21% (p=0.000 n=8+9)
name old allocs/op new allocs/op delta
Cache-12 5.00 ± 0% 2.00 ± 0% -60.00% (p=0.000 n=10+10)
$ benchstat striped.log expiring.log
name old time/op new time/op delta
Cache-12 1.58µs ± 5% 2.02µs ± 3% +27.34% (p=0.000 n=10+9)
name old alloc/op new alloc/op delta
Cache-12 288B ± 0% 107B ± 4% -62.85% (p=0.000 n=10+9)
name old allocs/op new allocs/op delta
Cache-12 9.00 ± 0% 2.00 ± 0% -77.78% (p=0.000 n=10+10)
$ benchstat simple.log striped.log expiring.log
name \ time/op simple.log striped.log expiring.log
Cache-12 2.74µs ± 2% 1.58µs ± 5% 2.02µs ± 3%
name \ alloc/op simple.log striped.log expiring.log
Cache-12 182B ± 0% 288B ± 0% 107B ± 4%
name \ allocs/op simple.log striped.log expiring.log
Cache-12 5.00 ± 0% 9.00 ± 0% 2.00 ± 0%
```
I also naively replacemed the LRU cache with the expiring cache in the
more realisitc CachedTokenAuthenticator benchmarks:
https://gist.github.com/mikedanese/41192b6eb62106c0758a4f4885bdad53
For token counts that fit in the LRU, expiring cache does better because
it does not require acquiring an exclusive lock for cache hits.
For token counts that exceed the size of the LRU, the LRU has a massive
performance drop off. The LRU cache is around 5x slower (with lookups
taking 1 milisecond and throttled to max 40 lookups in flight).
```
$ benchstat before.log after.log
name old time/op new time/op delta
CachedTokenAuthenticator/tokens=100_threads=256-12 3.60µs ±22% 1.08µs ± 4% -69.91% (p=0.000 n=10+10)
CachedTokenAuthenticator/tokens=500_threads=256-12 3.94µs ±19% 1.20µs ± 3% -69.57% (p=0.000 n=10+10)
CachedTokenAuthenticator/tokens=2500_threads=256-12 3.07µs ± 6% 1.17µs ± 1% -61.87% (p=0.000 n=9+10)
CachedTokenAuthenticator/tokens=12500_threads=256-12 3.16µs ±17% 1.38µs ± 1% -56.23% (p=0.000 n=10+10)
CachedTokenAuthenticator/tokens=62500_threads=256-12 15.0µs ± 1% 2.9µs ± 3% -80.71% (p=0.000 n=10+10)
name old alloc/op new alloc/op delta
CachedTokenAuthenticator/tokens=100_threads=256-12 337B ± 1% 300B ± 0% -11.06% (p=0.000 n=10+8)
CachedTokenAuthenticator/tokens=500_threads=256-12 307B ± 1% 304B ± 0% -0.96% (p=0.000 n=9+10)
CachedTokenAuthenticator/tokens=2500_threads=256-12 337B ± 1% 304B ± 0% -9.79% (p=0.000 n=10+10)
CachedTokenAuthenticator/tokens=12500_threads=256-12 343B ± 1% 276B ± 0% -19.58% (p=0.000 n=10+10)
CachedTokenAuthenticator/tokens=62500_threads=256-12 493B ± 0% 334B ± 0% -32.12% (p=0.000 n=10+10)
name old allocs/op new allocs/op delta
CachedTokenAuthenticator/tokens=100_threads=256-12 13.0 ± 0% 11.0 ± 0% -15.38% (p=0.000 n=10+10)
CachedTokenAuthenticator/tokens=500_threads=256-12 12.0 ± 0% 11.0 ± 0% -8.33% (p=0.000 n=10+10)
CachedTokenAuthenticator/tokens=2500_threads=256-12 13.0 ± 0% 11.0 ± 0% -15.38% (p=0.000 n=10+10)
CachedTokenAuthenticator/tokens=12500_threads=256-12 13.0 ± 0% 10.0 ± 0% -23.08% (p=0.000 n=9+10)
CachedTokenAuthenticator/tokens=62500_threads=256-12 17.0 ± 0% 12.0 ± 0% -29.41% (p=0.000 n=10+10)
```
Benchmarked with changes in #84423
Bugs: #83259#83375
(1) Replaced random-looking assortment of counter increments and
decrements with something hopefully more principalled-looking. Most
importantly, introduced the MutablePromise abstraction to neatly wrap
up the complicated business of unioning multiple sources of
unblocking.
(2) Improved debug logging.
(3) Somewhat more interesting test cases, and a bug fix wrt round
robin index.
Sometimes, volume creation can succeed right as the request times out,
causing k8s to interpret it as a failure. When the request is retried,
we want it to succeed. When trying this in vSphere, the second create
request failed with "already exists" and it never recovered.
This adds a check to the in-tree vsphere storage plugin that checks if a
VMDK exists before trying to create it. The check is done BEFORE create.
Tested: manual only )-:
