[
https://issues.apache.org/jira/browse/HDFS-12618?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16220685#comment-16220685
]
Daryn Sharp commented on HDFS-12618:
------------------------------------
bq. I did not -1 on this one because this only happens when someone runs
-includeSnapshots explicitly. Not sure if these snapshot problems can be solved
without doing this, please feel free to share any alternatives in mind.
Been awhile since I (attempted to) analyze snapshots. I think snapshotted
files, even deleted ones, are a {{INodeReference.WithName}} with a parent of
{{INodeReference.WithCount}} which maintains a list of all
{{INodeReference.WithName}}. Perhaps we could detect whether the inode is
linked into the current namesystem. If yes, it will be picked up in the
namesystem crawl; if no, count it based on all the {{WithName}} refs. And/or
maybe count a reference only if it's the last ref
({{INodeReference.WithCount#getLastWithName}}). Maybe.
bq. For large clusters doing fsck alone on / are a bad idea.
We do this nightly. Every day. Every cluster.
bq. Would it work for you if we put a memory limit on how much each fsck
-includeSnapshots' block map could consume on the NN?
I'm not sure how that could work in a user-friendly manner. I run the fsck, it
fails. I have to run it again on subdirs? Some fail again. I have to run it
on lower subdirs, then write code to collate all the mini-reports back together
in a unified report?
Fsck can run for tens of minutes or hours. Keeping excessively large state
during the operation will cause lots of pressure on old and risk OOM. It has
to stay lightweight (or only as heavy as it already is).
bq. Given this check is just relevant for files that had been deleted and
reside on snapshots only, would it still be a possibility for these files to be
truncated/appended?
I'm by no means a truncate/append + snapshot expert. A renamed file appears as
a delete in a snapshot diff. It can be subsequent modified in later versions
that may also be snapshotted and "deleted". Correct snapshot handling has been
problematic so we need to ensure it works correctly in all cases w/o causing
significant issues.
> fsck -includeSnapshots reports wrong amount of total blocks
> -----------------------------------------------------------
>
> Key: HDFS-12618
> URL: https://issues.apache.org/jira/browse/HDFS-12618
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: tools
> Affects Versions: 3.0.0-alpha3
> Reporter: Wellington Chevreuil
> Assignee: Wellington Chevreuil
> Priority: Minor
> Attachments: HDFS-121618.initial, HDFS-12618.001.patch,
> HDFS-12618.002.patch
>
>
> When snapshot is enabled, if a file is deleted but is contained by a
> snapshot, *fsck* will not reported blocks for such file, showing different
> number of *total blocks* than what is exposed in the Web UI.
> This should be fine, as *fsck* provides *-includeSnapshots* option. The
> problem is that *-includeSnapshots* option causes *fsck* to count blocks for
> every occurrence of a file on snapshots, which is wrong because these blocks
> should be counted only once (for instance, if a 100MB file is present on 3
> snapshots, it would still map to one block only in hdfs). This causes fsck to
> report much more blocks than what actually exist in hdfs and is reported in
> the Web UI.
> Here's an example:
> 1) HDFS has two files of 2 blocks each:
> {noformat}
> $ hdfs dfs -ls -R /
> drwxr-xr-x - root supergroup 0 2017-10-07 21:21 /snap-test
> -rw-r--r-- 1 root supergroup 209715200 2017-10-07 20:16 /snap-test/file1
> -rw-r--r-- 1 root supergroup 209715200 2017-10-07 20:17 /snap-test/file2
> drwxr-xr-x - root supergroup 0 2017-05-13 13:03 /test
> {noformat}
> 2) There are two snapshots, with the two files present on each of the
> snapshots:
> {noformat}
> $ hdfs dfs -ls -R /snap-test/.snapshot
> drwxr-xr-x - root supergroup 0 2017-10-07 21:21
> /snap-test/.snapshot/snap1
> -rw-r--r-- 1 root supergroup 209715200 2017-10-07 20:16
> /snap-test/.snapshot/snap1/file1
> -rw-r--r-- 1 root supergroup 209715200 2017-10-07 20:17
> /snap-test/.snapshot/snap1/file2
> drwxr-xr-x - root supergroup 0 2017-10-07 21:21
> /snap-test/.snapshot/snap2
> -rw-r--r-- 1 root supergroup 209715200 2017-10-07 20:16
> /snap-test/.snapshot/snap2/file1
> -rw-r--r-- 1 root supergroup 209715200 2017-10-07 20:17
> /snap-test/.snapshot/snap2/file2
> {noformat}
> 3) *fsck -includeSnapshots* reports 12 blocks in total (4 blocks for the
> normal file path, plus 4 blocks for each snapshot path):
> {noformat}
> $ hdfs fsck / -includeSnapshots
> FSCK started by root (auth:SIMPLE) from /127.0.0.1 for path / at Mon Oct 09
> 15:15:36 BST 2017
> Status: HEALTHY
> Number of data-nodes: 1
> Number of racks: 1
> Total dirs: 6
> Total symlinks: 0
> Replicated Blocks:
> Total size: 1258291200 B
> Total files: 6
> Total blocks (validated): 12 (avg. block size 104857600 B)
> Minimally replicated blocks: 12 (100.0 %)
> Over-replicated blocks: 0 (0.0 %)
> Under-replicated blocks: 0 (0.0 %)
> Mis-replicated blocks: 0 (0.0 %)
> Default replication factor: 1
> Average block replication: 1.0
> Missing blocks: 0
> Corrupt blocks: 0
> Missing replicas: 0 (0.0 %)
> {noformat}
> 4) Web UI shows the correct number (4 blocks only):
> {noformat}
> Security is off.
> Safemode is off.
> 5 files and directories, 4 blocks = 9 total filesystem object(s).
> {noformat}
> I would like to work on this solution, will propose an initial solution
> shortly.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
