Re: IO related waits - Mailing list pgsql-general

On 9/16/24 20:55, veem v wrote:
> 
> 
> On Tue, 17 Sept 2024 at 03:41, Adrian Klaver <adrian.klaver@aklaver.com 
> <mailto:adrian.klaver@aklaver.com>> wrote:
> 
> 
>     Are you referring to this?:
> 
>     https://nightlies.apache.org/flink/flink-docs-release-1.20/docs/dev/datastream/operators/asyncio/
<https://nightlies.apache.org/flink/flink-docs-release-1.20/docs/dev/datastream/operators/asyncio/>
> 
>     If not then you will need to be more specific.
> 
> 
> Yes, I was referring to this one. So what can be the caveats in this 
> approach, considering transactions meant to be ACID compliant as 
> financial transactions.Additionally I was not aware of the parameter 
> "synchronous_commit" in DB side which will mimic the synchronous commit.
> 
> Would both of these mimic the same asynchronous behaviour and achieves 
> the same, which means the client data load throughput will increase 
> because the DB will not wait for those data to be written to the WAL and 
> give a confirmation back to the client and also the client will not wait 
> for the DB to give a confirmation back on the data to be persisted in 
> the DB or not?. Also, as in the backend the flushing of the WAL to the 
> disk has to happen anyway(just that it will be delayed now), so can this 
> method cause contention in the database storage side if the speed in 
> which the data gets ingested from the client is not getting written to 
> the disk , and if it can someway impact the data consistency for the 
> read queries?

This is not something that I am that familiar with. I suspect though 
this is more complicated then you think. From the link above:

" Prerequisites #

As illustrated in the section above, implementing proper asynchronous 
I/O to a database (or key/value store) requires a client to that 
database that supports asynchronous requests. Many popular databases 
offer such a client.

In the absence of such a client, one can try and turn a synchronous 
client into a limited concurrent client by creating multiple clients and 
handling the synchronous calls with a thread pool. However, this 
approach is usually less efficient than a proper asynchronous client.
"

Which means you need to on Flink end:

1) Use Flink async I/O .

2) Find a client that supports async or fake it by using multiple 
synchronous clients.

On Postgres end there is this:

https://www.postgresql.org/docs/current/wal-async-commit.html

That will return a success signal to the client quicker if 
synchronous_commit is set to off. Though the point of the Flink async 
I/O is not to wait for the response before moving on, so I am not sure 
how much synchronous_commit = off would help.

-- 
Adrian Klaver
adrian.klaver@aklaver.com