substantially higher CPU after upgrading to v9.17.0

[clambin]

Issue tracker is used for reporting bugs and discussing new features. Please use
stackoverflow for supporting issues.

Since I upgraded my application to v9.17.0, I've seen a substantial increase in CPU usage: whereas previously the app was using no CPU when dormant, after the upgrade it's consistently using 10% CPU.

The only difference is an upgrade from v9.16.0 to v9.17.0.

Expected Behavior

Would expect same resource usage as on v9.16.0.

Current Behavior

CPU is much higher, even when dormant.

Steps to Reproduce

My app creates a simple client as follows:

		client := redis.NewClient(&redis.Options{
			Addr:     configuration.Redis.Addr,
			Username: configuration.Redis.Username,
			Password: configuration.Redis.Password,
			DB:       configuration.Redis.Database,
		})

Running pprof, I see a 60% of time spent in syscall.Syscall6, triggered by syscall.EpollWait (50%), called by runtime.findRunnable. So, it looks like something is constantly waking up the scheduler.

Context (Environment)

• Go 1.25.4
• Runtime: linux 6.8.0-1040-raspi running on arm64

[ndyakov]

Hello @clambin! Thank you for the investigation. Sounds like a busy loop (which when in use, should be better). I did not test the idle cpu usage, only when there is some traffic. Can you check if you have higher cpu usage when in use again?

[ndyakov]

Hello @clambin I found something that may be the reason, we did an optimization of a hot path that included a background precalculation of things that are needed in the hot path so they are not calculated on execution, but maybe this results in higher cpu usage. For this reason and the reason that I am not convinced by the benefit of this precalculation, I will open a PR with the "optimization" reverted and provide it to you for testing.

[ndyakov]

@clambin when you have the time, please test with #3611 if it resolves the cpu usage increase you are observing. Keep in mind there may be other things (background reconnects, etc) that may show as increased cpu usage as well, but their impact should not be that significant.

[clambin]

that's much better!

this is my test code:

func idle(ctx context.Context, c *redis.Client, duration time.Duration) error {
	ctx, cancel := context.WithTimeout(ctx, duration)
	defer cancel()
	for {
		select {
		case <-ctx.Done():
			return nil
		case <-time.After(5 * time.Second):
			if err := c.Ping(ctx).Err(); err != nil {
				if errors.Is(err, context.DeadlineExceeded) {
					err = nil
				}
				return err
			}
		}
	}
}

I ran this against 9.16, 9.17 and #3611 (pprof output attached):

v9.16:
real    1m 0.02s
user    0m 0.01s
sys     0m 0.04s

v9.17:
real    1m 0.02s
user    0m 0.13s
sys     0m 0.39s

patch:
real    1m 0.02s
user    0m 0.01s
sys     0m 0.05s

So, the revised version brings performance back to v9.16 levels. Great job!

[ndyakov]

Thank you @clambin, we will address couple other reports related to 9.17 and will release 9.17.1 patch version.