[
https://issues.apache.org/jira/browse/HADOOP-11238?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14238830#comment-14238830
]
Colin Patrick McCabe commented on HADOOP-11238:
-----------------------------------------------
bq. It's currently like this because the test case uses dependency injection to
test timing. We could use a fake guava timer but I wanted to avoid tight
coupling between hadoop and the guava library. Let me know what you think.
Sure, seems reasonable.
bq. Good idea. I added {{\[expireAfterWrite\]}} using 10*cacheTimeout, and
added some comments
Great.
{code}
201 * This method will block if a cache entry doesn't exist, and
202 * any subsequent requests for the user will wait on the first
203 * request to return. If a user already exists in the cache,
204 * this will be run in the background.
{code}
Maybe say "any subsequent requests for the user will wait on *this* request to
return" to make it slightly clearer.
Small nit: can you give different names to your different patches? There was a
discussion about this on hadoop-common-dev recently, the consensus is that
putting numbers on each one makes it clearer which is the latest (yes, we know
JIRA attaches a date as well :)
I renamed the jira to "Update the NameNode's Group Cache in the background when
possible" to better reflect the current patch (i.e. talked about the solution,
as well as the problem)
Thanks again.
> Update the NameNode's Group Cache in the background when possible
> -----------------------------------------------------------------
>
> Key: HADOOP-11238
> URL: https://issues.apache.org/jira/browse/HADOOP-11238
> Project: Hadoop Common
> Issue Type: Bug
> Affects Versions: 2.5.1
> Reporter: Chris Li
> Assignee: Chris Li
> Priority: Minor
> Attachments: HADOOP-11238.patch, HADOOP-11238.patch
>
>
> This patch addresses an issue where the namenode pauses during group
> resolution by only allowing a single group resolution query on expiry. There
> are two scenarios:
> 1. When there is not yet a value in the cache, all threads which make a
> request will block while a single thread fetches the value.
> 2. When there is already a value in the cache and it is expired, the new
> value will be fetched in the background while the old value is used by other
> threads
> This is handled by guava's cache.
> Negative caching is a feature built into the groups cache, and since guava's
> caches don't support different expiration times, we have a separate negative
> cache which masks the guava cache: if an element exists in the negative cache
> and isn't expired, we return it.
> In total the logic for fetching a group is:
> 1. If username exists in static cache, return the value (this was already
> present)
> 2. If username exists in negative cache and negative cache is not expired,
> raise an exception as usual
> 3. Otherwise Defer to guava cache (see two scenarios above)
> Original Issue Below:
> ----------------------------
> Our namenode pauses for 12-60 seconds several times every hour. During these
> pauses, no new requests can come in.
> Around the time of pauses, we have log messages such as:
> 2014-10-22 13:24:22,688 WARN org.apache.hadoop.security.Groups: Potential
> performance problem: getGroups(user=xxxxx) took 34507 milliseconds.
> The current theory is:
> 1. Groups has a cache that is refreshed periodically. Each entry has a cache
> expiry.
> 2. When a cache entry expires, multiple threads can see this expiration and
> then we have a thundering herd effect where all these threads hit the wire
> and overwhelm our LDAP servers (we are using ShellBasedUnixGroupsMapping with
> sssd, how this happens has yet to be established)
> 3. group resolution queries begin to take longer, I've observed it taking 1.2
> seconds instead of the usual 0.01-0.03 seconds when measuring in the shell
> `time groups myself`
> 4. If there is mutual exclusion somewhere along this path, a 1 second pause
> could lead to a 60 second pause as all the threads compete for the resource.
> The exact cause hasn't been established
> Potential solutions include:
> 1. Increasing group cache time, which will make the issue less frequent
> 2. Rolling evictions of the cache so we prevent the large spike in LDAP
> queries
> 3. Gate the cache refresh so that only one thread is responsible for
> refreshing the cache
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
