Behavior in event of disk i/o errors?

[jumbo jim <ju...@gmail.com>]

Cluster of unreliable commodity hardware.

Given the nature of master/master replication, I would imagine that you
could run cluster nodes without any form of raid?

As document commits are "all or nothing", I would assume that auto
replication would not occur if io errors were experienced.

How might couchdb behave in the event of io errors during reads?

Re: Behavior in event of disk i/o errors?

[Jan Lehnardt <ja...@apache.org>]

> On 01 Jan 2015, at 12:30 , jumbo jim <ju...@gmail.com> wrote:
> 
> Cluster of unreliable commodity hardware.
> 
> Given the nature of master/master replication, I would imagine that you
> could run cluster nodes without any form of raid?

This is a calculation of odds: how likely is it that your node fails vs.
how likely is it that raid fails vs. how likely it is that a single disk
fails, vs. how expensive is it to run either of these. All of this depends
on your setup, data and budget, but your suggestions might fit your setup.

> As document commits are "all or nothing", I would assume that auto
> replication would not occur if io errors were experienced.
> 
> How might couchdb behave in the event of io errors during reads?

we checksum data on disk and throw errors if the underlying blocks have
changed.

E.g. if you have an i/o error (of the nature that a block on disk was
flipped behind the back of CouchDB), that error will not replicate to
any replication targets.

Best
Jan
--