geode-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Kenneth Howe (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (GEODE-5513) Clients may miss PR region events due to race during registerInterest
Date Wed, 01 Aug 2018 16:56:00 GMT

     [ https://issues.apache.org/jira/browse/GEODE-5513?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Kenneth Howe updated GEODE-5513:
--------------------------------
    Affects Version/s: 1.6.0

> Clients may miss PR region events due to race during registerInterest
> ---------------------------------------------------------------------
>
>                 Key: GEODE-5513
>                 URL: https://issues.apache.org/jira/browse/GEODE-5513
>             Project: Geode
>          Issue Type: Bug
>          Components: client queues
>    Affects Versions: 1.6.0
>            Reporter: Kenneth Howe
>            Assignee: Kenneth Howe
>            Priority: Major
>
> Here is the scenario:
>  Consider two servers and client:
>  - Server1 hosting the primary bucket
>  - Server2 hosting secondary bucket and also primary queue for the Client2
>  - Client1 Doing remove operation
>  - Client2 doing register interest
> - The Client1 starts remove-all operation
>  - At the same time Client2 is registering interest
>  - Server1 receives the remove-all operation processes it, and sends the adjunct message
to the Server2 (Its still not yet received the interest info from server1)
>  - While the remove-all to server2 in flight
>  - Server2 sends interest profile info to Server1 for client2; and then Server2 (as it
is hosting the primary queue) starts building the initial image snapshot for the interest.
When building initial image for PR preference is given to collect data from local node. During
this time the removal message is still in flight and hasn't applied on Server2. The initial
image for interest registration calculates the snapshot from local data, and sends it to client,
missing the remove-all op.
> This could happen with non-bulk ops; but it gets worse with bulk ops as the time taken
to replicate the bulk ops will take more time.
> The solution is to build the initial register interest response by getting the data from
primary bucket. This will add little overhead in building the interest response; but considering
that most or always the register response will involve remote node, this may be negligible.
> Clients registering interest in a region



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Mime
View raw message