Colin, Thanks for the response, understanding the details is important and I think some general guidelines would be great. Since my initial email the system administrators told me that the drives are not actually full; the filesystems by default keep 5% in reserve. We can lower the reserve by 1%, which frees up about 30GB on a 3TB drive. We have been able to do just that, restart the DNs, and then everything reported in. At this point I can run the balancer and try to even things out across the DNs. Running the balancer before the upgrade was not really an option since in CDH 5.1.2 it did not work due to HDFS-6621. As a side note, I did apply the fix from HDFS-6621 and the balancer still didn't work (or I did someting wrong).
I think adding some properties for the upgrade would be a good idea. I would rather see an upgrade fail due to "not enough space" than have an upgrade appear to succeed and then end up with missing blocks. From a user perspective, the upgrade just failed in a bad way. Regarding existing failed drives, maybe the upgrade instructions should have the system administrators unmount them first. Dave Marion ----- Original Message ----- From: "Colin P. McCabe" <cmcc...@apache.org> To: hdfs-dev@hadoop.apache.org Sent: Thursday, January 8, 2015 2:43:29 PM Subject: Re: HDFS 2.6.0 upgrade ends with missing blocks Hi dlmarion, In general, any upgrade process we do will consume disk space, because it's creating hardlinks and a new "current" directory, and so forth. So upgrading when disk space is very low is a bad idea in any scenario. It's certainly a good idea to free up some space before doing the upgrade. We should put a note about this somewhere. We should also add this check into various Hadoop management software packages. Unfortunately, I can't tell you how much space you need exactly to do an upgrade. As you guessed, it depends on the local UNIX filesystem you are using. It also depends on how you formatted the filesystem. If you are using ext4, this parameter is controlled by the bytes-per-inode of the filesystem, multiplied by the number of inodes. ext4 may also run out of inodes even though there is free space on the disk (but this is very unlikely when using HDFS with reasonably large block sizes.) What I would say is that you need to have at least as many free inodes as blocks, and also a few kilobytes of space for writing some new VERSION files and other metadata files. In general, I would say leave a generous number of megabytes free to avoid problems with things like the block scanner filling up your disk unexpectedly, or the rebalancer filling up your disk during an MR job or other job. You mentioned that the DataNode continued to start even though some of the volumes couldn't be upgraded, due to dfs.datanode.failed.volumes.tolerated being non-zero. That is definitely unfortunate. Maybe we could add another configuration named dfs.datanode.failed.upgrade.volumes.tolerated, that took effect when upgrading. That way you could specify that any failures were not acceptable during an upgrade... i.e., they caused the upgrade to fail. Would this be helpful? The main downside that I can see for this-- and it is a big one-- is that if there are any existing failed drives, then the upgrade would not succeed. I think this would be scary for a lot of administrators. best, Colin Cloudera On Wed, Jan 7, 2015 at 5:25 AM, <dlmar...@comcast.net> wrote: > > I recently upgraded from CDH 5.1.2 to CDH 5.3.0. I know, contact Cloudera, > but this is actually a generic issue. After the upgrade I brought up the DNs > and after all of them had checked in I ended up with missing blocks. I > tracked this down in the DN logs to an error at startup where the DN is > failing to create subdirectories. This happens at > BlockPoolSliceStorage.doUpgrade(). It appears that the directory structure > has changed with HDFS-6482 and the DN is pre-creating all of the directories > at DN startup time. If the disk is near full, then it fails to create the > subdirectories because it consumes the remaining space. If the hdfs > configuration allows failed drives (dfs.datanode.failed.volumes.tolerated > > 0), then the DN will start without the now full disk and report all of the > blocks except the ones on the full disk. > > I didn't find any type of warning in the Apache release notes. It might be > useful for people in a similar situation. For the Cloudera folks on this > list, there is no warning or note in your upgrade instructions that I could > find either. > > Some questions: > > 1. How much free space is needed per disk to pre-create the directory > structure. Is it dependent on the type of filesystem? I calculated 256MB > given my reading of the ticket, but I may have misunderstood something. > > 2. Now that block locations are calculated using the block id, are there > restrictions on where blocks can be placed? I assume that the location is not > verified on a read for backwards compatibility, if that is not true, then > someone needs to comment on HDFS-1312 that the older utilities cannot be > used. I need to move blocks from the full disks to other locations, I'm > looking for any restrictions in doing that.
