[
https://issues.apache.org/jira/browse/SENTRY-893?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14905198#comment-14905198
]
Sravya Tirukkovalur edited comment on SENTRY-893 at 9/23/15 8:47 PM:
---------------------------------------------------------------------
I made a few more changes.
Sentry client is an instance variable of SimpleDBProviderBackend and hence it
is still possible that a thread is trying to access a sentry client with
previously broken/misaligned socket as the new thread is just blocked on the
synchronized call but still has a handle for the old client. For this reason I
think it is best to recreate a sentry client for each request. Updated the
patch to achieve thread safety by making sentry client a local variable. This
should not be a problem with pooled connection factory as we invalidate the
broken clients from the pool and provide only good clients.
was (Author: sravya):
I made a few more changes.
Right now, there is just one sentry client per HMS instance as (Sentry Client):
(SimpleDBProviderBackend) : (PolicyEngine): (AuthorizationProvider):
(HiveAuthzBinding): (MetaStoreAuthzBinding) = 1:1:1:1:1. And Sentry client is
an instance variable of SimpleDBProviderBackend and hence it is still possible
that a thread is trying to access a sentry client with previously
broken/misaligned socket as the new thread is just blocked on the synchronized
call but still has a handle for the old client. For this reason I think it is
best to recreate a sentry client for each request. Updated the patch to achieve
thread safety by making sentry client a local variable. This should not be a
problem with pooled connection factory as we invalidate the broken clients from
the pool and provide only good clients.
> Synchronize calls in SentryClient
> ---------------------------------
>
> Key: SENTRY-893
> URL: https://issues.apache.org/jira/browse/SENTRY-893
> Project: Sentry
> Issue Type: Bug
> Affects Versions: 1.4.0
> Reporter: Sravya Tirukkovalur
> Assignee: Sravya Tirukkovalur
> Fix For: 1.7.0
>
> Attachments: SENTRY-893.1.patch, SENTRY-893.patch
>
>
> We periodically see "out of sequence response" error from thrift layer as
> thrift itself is not thread safe and lets mutliple clients read each other's
> sockets. See
> http://mail-archives.apache.org/mod_mbox/thrift-user/201208.mbox/%3ca0f963dcf29346458cdf2969683df6cc70f90...@sc-mbx01-2.thefacebook.com%3E
> bq. Caused by: sentry.org.apache.thrift.TApplicationException:
> list_sentry_privileges_for_provider failed: out of sequence response at
> sentry.org.apache.thrift.TServiceClient.receiveBase(TServiceClient.java:76)
> Would be best to synchronize the calls on client side to avoid this
> situation. We might follow up to allow locking just the transport.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
