[
https://issues.apache.org/jira/browse/YARN-5167?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15303275#comment-15303275
]
Joep Rottinghuis commented on YARN-5167:
----------------------------------------
As part of our discussion of cost of doing all these replace operations we said
the keys shouldn't be _that_ long, so we shouldn't worry about the cost too
much unless we can show that it is a problem.
That thought lead us to the realization that there are currently no limits on
size of keys, nor on the size of the value passed to us.
By default HBase will allow a keyvalue size (configurable through
hbase.client.keyvalue.maxsize) of 10MB.
We said we should probably limit the keys to be no larger than a thousand
characters or so. The rowkey and column qualifiers would together get to a
considerable size, tags could be in the mix as well.
In order to avoid issues with region servers OOM'ing (on coprocessors),
replication between HBase clusters in different DCs choking, and clients dying
with memory issues we should probably enforce a reasonable limit.
We can make this configurable, but if we choose something like 2048 in max
rowkey, and column qualifier size and 127 MB for the value, then we can arrive
at a total max keyvalue size of max 128MB. [~vrushalic] will file a separate
jira for this and then we don't have to worry about walking through strings
that are too large in this jira.
> Escaping occurences of encodedValues
> ------------------------------------
>
> Key: YARN-5167
> URL: https://issues.apache.org/jira/browse/YARN-5167
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: timelineserver
> Reporter: Joep Rottinghuis
> Assignee: Sangjin Lee
> Priority: Critical
>
> We had earlier decided to punt on this, but in discussing YARN-5109 we
> thought it would be best to just be safe rather than sorry later on.
> Encoded sequences can occur in the original string, especially in case of
> "foreign key" if we decide to have lookups.
> For example, space is encoded as %2$.
> Encoding "String with %2$ in it" would decode to "String with in it".
> We though we should first escape existing occurrences of encoded strings by
> prefixing a backslash (even if there is already a backslash that should be
> ok). Then we should replace all unencoded strings.
> On the way out, we should replace all occurrences of our encoded string to
> the original except when it is prefixed by an escape character. Lastly we
> should strip off the one additional backslash in front of each remaining
> (escaped) sequence.
> If we add the following entry to TestSeparator#testEncodeDecode() that
> demonstrates what this jira should accomplish:
> {code}
> testEncodeDecode("Double-escape %2$ and %3$ or \\%2$ or \\%3$, nor
> \\\\%2$ = no problem!", Separator.QUALIFIERS,
> Separator.VALUES, Separator.SPACE, Separator.TAB);
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
