We have Varnish 3.0.2 running on Amazon's Linux and it works great. We have a ttl of 48 hours for most content pages and much longer for images, PDFs etc.

This weekend we've taken the backend down for some maintenance, so I upped the ttl to 5 days earlier in the week. I had assumed that anything in cache would continue to be served for up to 5 days, but much to our disappointment we checked varnishstat this morning and the cache was almost completely empty and varnish was serving "page not found" messages.

I know that this is not what Varnish is designed to do, but why would it reset its cache when the backend is down? And how can I prevent it for next time?

Update 2012-06-11: After looking in the /var/log/messages I see every 3 hours or so:

Jun  9 03:56:31 idea-varnish varnishd[1128]: Manager got SIGINT
Jun  9 03:56:33 idea-varnish varnishd[6708]: Platform: Linux,3.2.18-1.26.6.amzn1.x86_64,x86_64,-smalloc,-smalloc,-hcritbit
Jun  9 03:56:33 idea-varnish varnishd[6708]: child (6709) Started
Jun  9 03:56:33 idea-varnish varnishd[6708]: Child (6709) said Child starts

I guess this is the server crashing and wiping all the objects in memory. I have only just now installed the -debuginfo rpm but not sure that will actually show anything more.

I supposed we could have switched back to disk-based storage during the scheduled downtime? or would a crash like this wipe that anyway?

  • You could probably use grace to do this, rather than increasing your TTLs :) varnish-cache.org/docs/trunk/tutorial/… – Kirrus Jun 10 '12 at 14:47
  • Grace mode seems to be for when the backend server is restarting or just temporarily overloaded. Our servers were scheduled to be down for 32+ hours, so I think that a 3d grace was a good idea since all the examples are under 5 minutes. – Nic Cottrell Jun 12 '12 at 8:01

Did the varnish process maybe restart? There's an uptime counter in varnishstat. Under certain circumstances the varnish worker thread can die, but it gets restarted immediately. When everything is working fine, this might go unnoticed, but with (planned) backend down time it can be quite inconvient.

  • I've since restarted manually, but will look out for this counter for next time. We've been watching varnishstat regularly and it always seems at about 99% hitrate, but maybe we've just never caught it after a crash... – Nic Cottrell Jun 11 '12 at 7:52
  • Yeah - it looks like something overloaded due to the backend being down so long. I believe that the newest 3.x release is more stable. – Nic Cottrell Oct 25 '12 at 8:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.