[
https://issues.apache.org/jira/browse/IGNITE-20934?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Vladimir Pligin updated IGNITE-20934:
-------------------------------------
Description:
At the moment, when executing an RW transaction, we will write to existing
(available and registered) indexes and to all dropped available indexes that
were in the past. This is necessary so as not to abort an RW transactions that
caught the index dropping event.
It won’t be good if we write to old dropped available indexes for a very long
time; we need to stop writing to them as quickly as possible. At the same time,
long RW transactions that saw the dropping of the index should not be broken.
We can get stuck on the mechanism of a low watermark, but if it is made
infinite, it will bring us a lot of pain.
You need to think carefully about solving the problem. I have the following
thoughts. Locally, each node will monitor the progress of its an RW
transactions and as soon as the last RW transaction that caught the index
dropping event disappears, immediately remove it from the selection for new RW
transactions.
was:
At the moment, when executing an RW transaction, we will write to existing
(available and registered) indexes and to all dropped available indexes that
were in the past. This is necessary so as not to abort an RW transactions that
caught the index dropping event.
It won’t be good if we write to old dropped available indexes for a very long
time; we need to stop writing to them as quickly as possible. At the same time,
long RW transactions that saw the dropping of the index should not be broken.
We can get stuck on the mechanism of a low watermark, but if it is made
infinite, it will bring us a lot of pain.
You need to think carefully about solving the problem. I have the following
thoughts. Locally, each node will monitor the progress of its an RW
transactions and as soon as the last RW transaction that caught the index
dropping event disappears, immediately remove it from the selection for new RW
transactions.
> Worry about deleted indexes that will no longer be used until they are
> destroyed
> --------------------------------------------------------------------------------
>
> Key: IGNITE-20934
> URL: https://issues.apache.org/jira/browse/IGNITE-20934
> Project: Ignite
> Issue Type: Improvement
> Reporter: Kirill Tkalenko
> Priority: Major
> Labels: ignite-3
>
> At the moment, when executing an RW transaction, we will write to existing
> (available and registered) indexes and to all dropped available indexes that
> were in the past. This is necessary so as not to abort an RW transactions
> that caught the index dropping event.
> It won’t be good if we write to old dropped available indexes for a very long
> time; we need to stop writing to them as quickly as possible. At the same
> time, long RW transactions that saw the dropping of the index should not be
> broken.
> We can get stuck on the mechanism of a low watermark, but if it is made
> infinite, it will bring us a lot of pain.
> You need to think carefully about solving the problem. I have the following
> thoughts. Locally, each node will monitor the progress of its an RW
> transactions and as soon as the last RW transaction that caught the index
> dropping event disappears, immediately remove it from the selection for new
> RW transactions.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
