Can errors occur while committing a tx?

Can an error occur /during/ the process of committing a transaction?
BeginTraction();
try {
UpdateTable1();
UpdateTable2();
CommitTransaction(); <-- error here?
} catch(Exception e){
RollbackTransaction();
}
Also, what are the implications in such a situation where both the
Updates pass without any errors, but committing the transaction fails?
Is this scenario even possible at all?
If so, what are the suggested best-practices for recovering from such
types of errors?
TIA,
Abdullah"Abdullah Kauchali" <none@.none.com> wrote in message
news:eT1t1UAzFHA.1264@.tk2msftngp13.phx.gbl...
> Can an error occur /during/ the process of committing a transaction?
> BeginTraction();
> try {
> UpdateTable1();
> UpdateTable2();
> CommitTransaction(); <-- error here?
> } catch(Exception e){
> RollbackTransaction();
throw;
> }
>
Remember to rethrow the exception!
You know, that's a really good question.
Yes errors are possible, but pretty darn unlikely (at least in SQL Server).
When you go to commit the transaction all of the changes have allready been
made to the tables, and written to the memory cache of the log file. So
almost everything that could go wrong already would have. There could
possibly be some error flushing the log to disk, or you could loose your
connection to the database server, or the server could just fail. If the
commit fails on the server, the transaction will be rolled back (or at worst
it won't be there when the database recovers). But from a client there's
probably some possiblility that the commit succeeds, but a network problem
prevents you from learning about it.

> Also, what are the implications in such a situation where both the
> Updates pass without any errors, but committing the transaction fails?
The transaction will be rolled back.

> Is this scenario even possible at all?
> If so, what are the suggested best-practices for recovering from such
> types of errors?
>
Treat it like a server or network failure.
It's so unlikely to happen in the first place, and you are so unlikely to be
able to recover if it does, that I would just pretend like it's impossible.
Just pretend like it can't happen. Your responsibility is only to keep the
database in a logically consistent state. You've done that by coding your
transaction. You are not responsible for making sure the transaction
suceeds. That responsibility belongs to a higher context (ie a user or an
automated agent).
For most programs any kind of transaction retry is not worth the coding.
The code complexity and residual risk are just too great. Just propagate
the error out to the user or calling code and let them deal with it. You
just don't have the right context to deal with a server or network failures
in a meaningful way.
If you feel like you must deal with it, then propagate the exception out of
this method, and catch it at the level which knows how to retry the entire
transaction. Wait around for the instance to fail over to another server,
reconnect and issue the transaction again. If you know you are running
against a cluster with such high availability requirements that it the DBA's
must fail the instance over to perform routine maintenance then you can
expect this to happen, and you have no choice but to code around it. But in
any case the retry code does not belong in that method, but in an outer
controlling context.
David

Can errors occur while committing a tx?

Can an error occur /during/ the process of committing a transaction?
BeginTraction();
try {
UpdateTable1();
UpdateTable2();
CommitTransaction(); <-- error here?
} catch(Exception e){
RollbackTransaction();
}
Also, what are the implications in such a situation where both the
Updates pass without any errors, but committing the transaction fails?
Is this scenario even possible at all?
If so, what are the suggested best-practices for recovering from such
types of errors?
TIA,
Abdullah"Abdullah Kauchali" <none@.none.com> wrote in message
news:eT1t1UAzFHA.1264@.tk2msftngp13.phx.gbl...
> Can an error occur /during/ the process of committing a transaction?
> BeginTraction();
> try {
> UpdateTable1();
> UpdateTable2();
> CommitTransaction(); <-- error here?
> } catch(Exception e){
> RollbackTransaction();
throw;
> }
>
Remember to rethrow the exception!
You know, that's a really good question.
Yes errors are possible, but pretty darn unlikely (at least in SQL Server).
When you go to commit the transaction all of the changes have allready been
made to the tables, and written to the memory cache of the log file. So
almost everything that could go wrong already would have. There could
possibly be some error flushing the log to disk, or you could loose your
connection to the database server, or the server could just fail. If the
commit fails on the server, the transaction will be rolled back (or at worst
it won't be there when the database recovers). But from a client there's
probably some possiblility that the commit succeeds, but a network problem
prevents you from learning about it.
> Also, what are the implications in such a situation where both the
> Updates pass without any errors, but committing the transaction fails?
The transaction will be rolled back.
> Is this scenario even possible at all?
> If so, what are the suggested best-practices for recovering from such
> types of errors?
>
Treat it like a server or network failure.
It's so unlikely to happen in the first place, and you are so unlikely to be
able to recover if it does, that I would just pretend like it's impossible.
Just pretend like it can't happen. Your responsibility is only to keep the
database in a logically consistent state. You've done that by coding your
transaction. You are not responsible for making sure the transaction
suceeds. That responsibility belongs to a higher context (ie a user or an
automated agent).
For most programs any kind of transaction retry is not worth the coding.
The code complexity and residual risk are just too great. Just propagate
the error out to the user or calling code and let them deal with it. You
just don't have the right context to deal with a server or network failures
in a meaningful way.
If you feel like you must deal with it, then propagate the exception out of
this method, and catch it at the level which knows how to retry the entire
transaction. Wait around for the instance to fail over to another server,
reconnect and issue the transaction again. If you know you are running
against a cluster with such high availability requirements that it the DBA's
must fail the instance over to perform routine maintenance then you can
expect this to happen, and you have no choice but to code around it. But in
any case the retry code does not belong in that method, but in an outer
controlling context.
David

Can errors occur while committing a tx?

Can an error occur /during/ the process of committing a transaction?
BeginTraction();
try {
UpdateTable1();
UpdateTable2();
CommitTransaction(); <-- error here?
} catch(Exception e){
RollbackTransaction();
}
Also, what are the implications in such a situation where both the
Updates pass without any errors, but committing the transaction fails?
Is this scenario even possible at all?
If so, what are the suggested best-practices for recovering from such
types of errors?
TIA,
Abdullah
"Abdullah Kauchali" <none@.none.com> wrote in message
news:eT1t1UAzFHA.1264@.tk2msftngp13.phx.gbl...
> Can an error occur /during/ the process of committing a transaction?
> BeginTraction();
> try {
> UpdateTable1();
> UpdateTable2();
> CommitTransaction(); <-- error here?
> } catch(Exception e){
> RollbackTransaction();
throw;
> }
>
Remember to rethrow the exception!
You know, that's a really good question.
Yes errors are possible, but pretty darn unlikely (at least in SQL Server).
When you go to commit the transaction all of the changes have allready been
made to the tables, and written to the memory cache of the log file. So
almost everything that could go wrong already would have. There could
possibly be some error flushing the log to disk, or you could loose your
connection to the database server, or the server could just fail. If the
commit fails on the server, the transaction will be rolled back (or at worst
it won't be there when the database recovers). But from a client there's
probably some possiblility that the commit succeeds, but a network problem
prevents you from learning about it.

> Also, what are the implications in such a situation where both the
> Updates pass without any errors, but committing the transaction fails?
The transaction will be rolled back.

> Is this scenario even possible at all?
> If so, what are the suggested best-practices for recovering from such
> types of errors?
>
Treat it like a server or network failure.
It's so unlikely to happen in the first place, and you are so unlikely to be
able to recover if it does, that I would just pretend like it's impossible.
Just pretend like it can't happen. Your responsibility is only to keep the
database in a logically consistent state. You've done that by coding your
transaction. You are not responsible for making sure the transaction
suceeds. That responsibility belongs to a higher context (ie a user or an
automated agent).
For most programs any kind of transaction retry is not worth the coding.
The code complexity and residual risk are just too great. Just propagate
the error out to the user or calling code and let them deal with it. You
just don't have the right context to deal with a server or network failures
in a meaningful way.
If you feel like you must deal with it, then propagate the exception out of
this method, and catch it at the level which knows how to retry the entire
transaction. Wait around for the instance to fail over to another server,
reconnect and issue the transaction again. If you know you are running
against a cluster with such high availability requirements that it the DBA's
must fail the instance over to perform routine maintenance then you can
expect this to happen, and you have no choice but to code around it. But in
any case the retry code does not belong in that method, but in an outer
controlling context.
David

Can a commit fail??

Can an error occur /during/ the process of committing a transaction? So:
BeginTraction();
try {
UpdateTable1();
UpdateTable2();
CommitTransaction(); <-- error here?
} catch(Exception e){
RollbackTransaction();
}
Also, what are the implications in such a situation where both the
Updates pass without any errors, but committing the transaction fails?
Is this scenario even possible at all?
If so, what are the suggested best-practices for recovering from such
types of errors?
TIA,
AbdullahYes. If the log disk fills up, the write of the commit sentinal can fail,
thus it's indeed possible for a commit to fail. I'm sure that there are
other scenarios that could cause a commit failure. Therefore, if the commit
fails, the transaction should be rolled back (if that hasn't already
happened as a result of the failure).
"Abdullah Kauchali" <none@.none.com> wrote in message
news:ukGjDWAzFHA.1264@.tk2msftngp13.phx.gbl...
> Can an error occur /during/ the process of committing a transaction? So:
> BeginTraction();
> try {
> UpdateTable1();
> UpdateTable2();
> CommitTransaction(); <-- error here?
> } catch(Exception e){
> RollbackTransaction();
> }
> Also, what are the implications in such a situation where both the
> Updates pass without any errors, but committing the transaction fails?
> Is this scenario even possible at all?
> If so, what are the suggested best-practices for recovering from such
> types of errors?
> TIA,
> Abdullah|||Brian Selzer wrote:
> Yes. If the log disk fills up, the write of the commit sentinal can fail,
> thus it's indeed possible for a commit to fail. I'm sure that there are
> other scenarios that could cause a commit failure. Therefore, if the comm
it
> fails, the transaction should be rolled back (if that hasn't already
> happened as a result of the failure).
Thanks Brian.
Let's use your example. Suppose UpdateTable1() succeeds during the
COMMIT (SQL Server frees the resources and releases the locks for
Table1) and then attempts to commit statements in UpdateTable2() but
then realises "oops, transaction log is full!" Will UpdateTable1()
rollback on a rollback instruction? Won't the transaction logs still
be considered full for the rollback log entries to go through?
:)
(I am actually trying to understand the process of commit in a 2-phase
scenario (distributed transactions), but I'd like to understand the
concept from a local-transaction point of view first. So, I apologise
for my lack of knowledge there!
My question for the 2-phase (distributed) transaction is this: if the
DTC commits all preceding resources and then encounters a problem with
the very last resource in the chain of updates, can/does the DTC
actually "uncommit" the preceding resources it just instructed to
commit? No during the prepare phase, but during the commit phase.)|||The commit of UpdateTable1() succeeded, so a rollback isn't possible. The
commit of UpdateTable2() fails, the log file is full, and the database shuts
down. During recovery (assuming disk space has been freed or otherwise made
available), UpdateTable2() will be rolled back. Changes are written to the
transaction log before they're written to the database, and only after all
of the database changes have been flushed to the disk is the commit sentinal
written to the transaction log, so the recovery process can undo any changes
made by any uncommitted transactions.
I'm not sure exactly how the process works with a distributed transaction.
Maybe there's a different type of sentinal written to the transaction log
after the prepare phase has completed. During the prepare phase, all cached
changes in each participant are flushed to the disk and then a
ready-to-commit signal is sent back to the coordinator. Once the commit
signal has been sent, I don't think a rollback is possible, even if an error
occurs on one of the other participants. If communication is lost before
the commit signal is received, then the participant is required to roll back
the transaction. If it happens afterward, the transaction is supposed to be
committed. Again, I'm not sure exactly how the process works under the
covers. Maybe someone with more knowledge than I can give you a more
difinitive answer.
"Abdullah Kauchali" <none@.none.com> wrote in message
news:uUupV7EzFHA.1856@.TK2MSFTNGP12.phx.gbl...
> Brian Selzer wrote:
> Thanks Brian.
> Let's use your example. Suppose UpdateTable1() succeeds during the
> COMMIT (SQL Server frees the resources and releases the locks for
> Table1) and then attempts to commit statements in UpdateTable2() but
> then realises "oops, transaction log is full!" Will UpdateTable1()
> rollback on a rollback instruction? Won't the transaction logs still
> be considered full for the rollback log entries to go through?
> :)
> (I am actually trying to understand the process of commit in a 2-phase
> scenario (distributed transactions), but I'd like to understand the
> concept from a local-transaction point of view first. So, I apologise
> for my lack of knowledge there!
> My question for the 2-phase (distributed) transaction is this: if the
> DTC commits all preceding resources and then encounters a problem with
> the very last resource in the chain of updates, can/does the DTC
> actually "uncommit" the preceding resources it just instructed to
> commit? No during the prepare phase, but during the commit phase.)
>