Thread: How to handle failed COMMIT

I've been wondering what the behavior of postgres is when the server process stops, for whatever reason, at certain
criticalpoints in the execution flow. 

In the following example the client will only regard the data as stored until the COMMIT command is successfully
executed.But the the server, client or network may fail at any point during the execution and therefore the server and
clientmay not be in sync of what the current state is. 

BEGIN;
INSERT INTO ....;
COMMIT;

To experiment with this I inserted a stupid if statement (see patch) which will make the server process exit(1) if the
clientsends a COMMIT command, but only after the COMMIT command has been processed on the server and just before the
serversend the close commend (wire protocol). I.e. the server has COMMITed the transaction, but the client just
experiencesthat the connection has been closed for some reason: 

server closed the connection unexpectedly
    This probably means the server terminated abnormally
    before or while processing the request.
The connection to the server was lost. Attempting reset: Succeeded.

So my question is, how should the client handle these situations? I think most systems will have some variation of the
following:

try
   execute transactional sql
catch (commit failed)
   // regard data as not stored

But, this doesn't seem to be enough, so I guess you would have to do something like:

try
   execute transactional sql
catch (commit failed)
   if (data is not stored)
     // regard data as not stored

Thanks in advance.

/Håvar Nøvik

try
   execute transactional sql
catch (commit failed)
   if (data is not stored)
     // regard data as not stored

On Monday, July 18, 2022, Håvar Nøvik <havar@novik.email> wrote:

try

   execute transactional sql

catch (commit failed)

   if (data is not stored)

     // regard data as not stored

Correct, the client did not get confirmation of commit success so it must operate as if it failed.

David J. 

> Correct, the client did not get confirmation of commit success so it must operate as if it failed.

I mean that’s the point, the client can’t operate as if it failed. It must operate as the state is unknown. But maybe that’s the correct application behaviour, just that I haven’t thought this through previously.