Re: bogus Directory Store Stats?

Henrik Nordstrom writes:

>Duane Wessels wrote:
>
>> I don't know what you mean by 'gone'? How does a cache_dir become gone?
>
>A cache_dir is effectively gone if the disk has failed, causing all
>writes to that particular cache_dir to fail.
>
>> If store_swap_size is 0 when a write() fails, then yes, it probably
>> reduces the size to zero.

Well, to clarify, its not just if write fails, but also
if errno == ENOSPC.

>What I ment was that if we only adjust the total, then a Squid running
>with one failed cache_dir may eventually end up with store_swap_size 0
>even thougth there is several other cache_dir directories that are
>helthy and fine. Well, it probably crashes before but ;-).
>
>The reason why I mention this is that I have a dream that it is possible
>to build a disk-fault tolerant cache server without relying on any type
>of RAID. If Squid could (gradually) phase out a disk that is bad without
>manual intervention then it is a big step on the way.

Sounds nice. How does the application decide the disk has failed?
10 write errors? 50% per 60 seconds?

I can imagine a situation where someone will have a failed disk and
never know it for three months because they don't look at the log
files and Squid just keeps on working. :-)

Duane W.
Received on Tue Jul 29 2003 - 13:15:51 MDT