[ 
https://issues.apache.org/jira/browse/IGNITE-27477?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Roman Puchkovskiy updated IGNITE-27477:
---------------------------------------
    Description: 
In IGNITE-27346, a problem of starting a node having stored LWM lower than the 
earliest Catalog timestamp was solved. The solution was to use Catalog's 
earliest timestamp instead of LWM (if LWM is lower) in each place where, during 
recovery, we want to access the Catalog history by LWM.

This seems to work now, but it creates a dangerous situation: some components 
are actually recovered with Catalog-rectified LWM, but we still leave the 
original LWM to be seen by other components (and permit it to remain after node 
start).

Another approach seems to be more reliable: that is, as soon as we recover the 
Catalog history, raise the LWM if needed. After that, all the components will 
see the rectified LWM in a consistent way, and its monotonicity will be 
maintained.

  was:
In IGNITE-27346, a problem of starting a node having stored LWM lower than the 
earliest Catalog timestamp was solved. The solution was to use Catalog's 
earliest timestamp instead of LWM (if LWM is lower) in each place where, during 
recovery, we want to access the Catalog history by LWM.

This


> Raise LWM on recovery if it lags behind earliest catalog
> --------------------------------------------------------
>
>                 Key: IGNITE-27477
>                 URL: https://issues.apache.org/jira/browse/IGNITE-27477
>             Project: Ignite
>          Issue Type: Improvement
>            Reporter: Roman Puchkovskiy
>            Assignee: Roman Puchkovskiy
>            Priority: Major
>              Labels: ignite-3
>
> In IGNITE-27346, a problem of starting a node having stored LWM lower than 
> the earliest Catalog timestamp was solved. The solution was to use Catalog's 
> earliest timestamp instead of LWM (if LWM is lower) in each place where, 
> during recovery, we want to access the Catalog history by LWM.
> This seems to work now, but it creates a dangerous situation: some components 
> are actually recovered with Catalog-rectified LWM, but we still leave the 
> original LWM to be seen by other components (and permit it to remain after 
> node start).
> Another approach seems to be more reliable: that is, as soon as we recover 
> the Catalog history, raise the LWM if needed. After that, all the components 
> will see the rectified LWM in a consistent way, and its monotonicity will be 
> maintained.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to