Google To Switch To EXT4, Hires Ted To Code

[Ant P.]

It might be useful to add in a latencytop benchmark too, to see whether it really is the cause of all the stuttering mouse pointer problems.

[bnolsen]

It might be fast, but it also has some pretty serious shortcomings (e.g. merging files from ReiserFS images into the filesystem itself). Reiser himself acknowledged this (his solution: ReiserFS v3 is obsolete, use Reiser4).

I'm not in any position to test filesystems anymore. I'm not sure if reiser4 actually ever got into a "stable" release?? Could be all the negative press. That being said, with ~7-8 years of use with at least 30 raid5/6's no raids were ever lost with reiser3. With much less use on a few choice systems, one raid was totally lost with xfs, and 2 were totally lost with jfs (like reformat/rebuild from scratch required). This is in a heavy imagery prcoessing shop where jobs can run 24/7 for days. And 24/7 uptime isn't an absolute requirement: catastrpphic power failure on the weekends or at night sometimes weren't "caught" in time (backhoes anyone?)

I'll float ext4 by the guy who's still there.

[deanjo]

I'm not in any position to test filesystems anymore. I'm not sure if reiser4 actually ever got into a "stable" release?? Could be all the negative press. That being said, with ~7-8 years of use with at least 30 raid5/6's no raids were ever lost with reiser3. With much less use on a few choice systems, one raid was totally lost with xfs, and 2 were totally lost with jfs (like reformat/rebuild from scratch required). This is in a heavy imagery prcoessing shop where jobs can run 24/7 for days. And 24/7 uptime isn't an absolute requirement: catastrpphic power failure on the weekends or at night sometimes weren't "caught" in time (backhoes anyone?)

I'll float ext4 by the guy who's still there.

Well anybody running a system that has critical data deserves to loose data if a proper UPS and monitor isn't setup and utilized. That's more of a IT failure then a file system failure IMHO and no file system can protect against administration stupidity.

[bnolsen]

Well anybody running a system that has critical data deserves to loose data if a proper UPS and monitor isn't setup and utilized. That's more of a IT failure then a file system failure IMHO and no file system can protect against administration stupidity.

UPSs don't last 12 hours, maybe 45 mins or so. And the people aren't paid to babysit this 24/7. Only 2 folks doing IT with several hundred terrabytes.

This isn't an IT failure its a cost vs availability decision. In the case of reiserfs its been proven in production there's no catastrophic risk for really bad power failures. With XFS and JFS i's proven to go catastrophic in really bad power fail conditions.

It means there is a working solution without having to buy a bank of diesel generators in case something really bad happens every year/couple of years.

[squirrl]

I've always used Reiser3 for my external drives and all file-systems.
Never have I once had any problems. I've used the file-system for more than 8 years.
I always shut my crap down the wrong way by hitting the power button on the power strip.


I don't want come off as sounding negative towards Theo Tso, as I'm sure there is always room for improvement in logic. Reiser himself stated he pulled an algorithm from a textbook in a moment of genious. Then there is that guy from down-under that wrote the brainfu.. scheduler. I think Google is better to stick with ext2 or move to Resierfs. Just depends on their set up. Ext3 would probably be a better choice given the data this site has turned up on ext4.

Now my negativity:
I have easily destroyed ext3 by plugging it into a usb-1.1 port and having something go haywire. It would no longer find the superblocks. I know about all the commands to restore one. Nothing worked. Reiser would quit reading and seg fault but disconnecting the drive and reconnection would always turn up my trusty archive of junk.

I also never was able to get the ext3 tool for win32 to read Ext4 partitions on the external in windows.


I'll assume they have a sort of Sans array set up at a bargain discount configuration and they are primarily implementing this for that. I read in the past they used Redhat for the core search hardware. So that probably has a lot to do with the decision.


L8r

[kernelOfTruth]

I think that is something that was actually significantly improved.

it also depends greatly on the amount of data currently on your partition:

if you have a 2 TB partition and only 20 GB on it - it of course would be much faster

take a look at the manpage then you'll know what I mean

[bnolsen]

it also depends greatly on the amount of data currently on your partition:

if you have a 2 TB partition and only 20 GB on it - it of course would be much faster

take a look at the manpage then you'll know what I mean

That would be either 10 or 14TB arrays, mostly between 2/3 and totally full. Stupid aerial imagery likes to multiply itself a lot.

Interesting comments about by squirt about ext3 reliability. Also interesting he's had similiar experiences in a very different domain.

Robustness is always a must. It's interesting both jfs and xfs both fail at this basic requirement.