On Apr 6, 2010, at 11:20 PM, Adam Kocoloski <[email protected]> wrote:

> On Apr 6, 2010, at 10:50 PM, Paul J Davis wrote:
> 
>> This corruption was quite odd in that there wasn't a conspicuous reason for 
>> it.  I didn't dive to deep into the whole thing so it's possible i missed 
>> something obvious. 
> 
> The instance was unresponsive to ssh for 12 hours.  The report from AWS 
> Support was merely a "problem with the underlying host" followed by a 
> recommendation to "launch a replacement at your earliest convenience".  I 
> don't know what the gremlins were doing behind the scenes, but I'm not 
> surprised the files are corrupted :)
> 

Yeah I don't think that we should worry about high energy particles flipping 
bits too much here.

>> There are two things at play here.  How proactive should we be in provoking 
>> theseI errors and how much should we check for situations where our data 
>> file got trounced.
>> 
>> The extreme proactive position would be equivalent to a full table scan per 
>> write which is out of the question. So to some extent we won't be able to 
>> detect some errors until read time which is an unknowable interval.
> 
> I'm totally comfortable with only detecting them at read-time.
> 
>> The other aspect is how rigorous should we check reads? This extreme would 
>> basically require a sha1 for every read or write no matter how small, not to 
>> mention the storage overhead. This part I'm not sure about. There's probably 
>> middle ground with crc sums and what not but i don't see a clear answer.
> 
> We currently store MD5 checksums with document bodies and validate them on 
> reads.  It hasn't proven to be an undue burden.
> 

We do that for every doc body? Did not know that. Perhaps general 
append_term_md5 usage wouldn't be as big of a deal as i feared.

> Best, Adam
> 
>> Basically, the question is how much should we attempt to detect when 
>> hardware lies.  I reckon that there's probably a middle ground to report 
>> when an assumption is violated and full on table scans. Ideally such things 
>> would be fairly configurable but i sure don't see an obvious answer.
>> 
>> 
>> On Apr 6, 2010, at 10:06 PM, Randall Leeds <[email protected]> wrote:
>> 
>>> I immediately want to say 'ini file option' but I'm not sure whether to err
>>> on safety or speed.
>>> 
>>> Maybe this is a good candidate for merkle trees or something else we can do
>>> throughout the view tree that might less overhead than md5 summing all the
>>> nodes? After all, most inner nodes shouldn't change most of the time. Some
>>> incremental, cheap checksum might be a worthwhile *option*.
>>> 
>>> On Apr 6, 2010 6:04 PM, "Adam Kocoloski" <[email protected]> wrote:
>>> 
>>> Hi all, we recently had an EC2 node go AWOL for about 12 hours.  When it
>>> came back, we noticed after a few days that a number of the view indexes
>>> stored on that node were not updating.  I did some digging into the error
>>> logs and with Paul's help pieced together what was going on.  I won't bother
>>> you with all the gory details unless you ask for them, but the gist of it is
>>> that those files are corrupted.
>>> 
>>> The troubling thing for me is that we only discovered the corruption when it
>>> completely broke the index updates.  In one case, it did this by rearranging
>>> the bits so that couch_file thought that the btree node it was reading from
>>> disk had an associated MD5 checksum. It didn't (no btree nodes do), and so
>>> couch_file threw a file_corruption exception.  But if the corruption had
>>> shown up in another part of the file I might never have known.  In fact,
>>> some of the other indices on that node probably are silently corrupted.
>>> 
>>> You might wonder how likely it is that a file becomes corrupted but still
>>> appears to be functioning.  I checked the last modified timestamps for three
>>> broken files.  One was last modified when the node went down, but the other
>>> two had timestamps in between the node's recovery and now.  To me, that
>>> means that the view indexer was able to update those files for quite a while
>>> (~2 days) before it bumped into a part of the btree that was corrupted.
>>> 
>>> I wonder what we should do about this.  My first thought is to make it
>>> optional to write  btree nodes (possibly only for view index files?) using
>>> append_term_md5 instead of append_term.  It seems like a simple patch, but I
>>> don't know a priori what the performance hit would be.  Other thoughts?
>>> 
>>> Best, Adam
> 

Reply via email to