Avoiding deadlocks ...

Started by Josh Berkusover 15 years ago11 messages
#1Josh Berkus
josh@agliodbs.com

Kevin,

This one is for you:

Two sessions, in transaction:

Process A Process B

update session where id = X;
update order where orderid = 5;
update order where orderid = 5;
update order where orderid = 5;
... deadlock error.

It seems like we ought to be able to avoid a deadlock in this case;
there's a clear precedence of who grabbed the order row first. Does
your serializability patch address the above situation at all?

--
-- Josh Berkus
PostgreSQL Experts Inc.
http://www.pgexperts.com

#2Josh Berkus
josh@agliodbs.com
In reply to: Josh Berkus (#1)
Re: Avoiding deadlocks ...

On 8/19/10 3:51 PM, Josh Berkus wrote:

Kevin,

This one is for you:

Two sessions, in transaction:

Process A Process B

update session where id = X;
update order where orderid = 5;
update order where orderid = 5;
update order where orderid = 5;
... deadlock error.

Johto on IRC pointed out I left something out of the above: "session" is
referenced in an FK by "orders", and session = X is related to orderid = 5.

It seems like we ought to be able to avoid a deadlock in this case;
there's a clear precedence of who grabbed the order row first. Does
your serializability patch address the above situation at all?

--
-- Josh Berkus
PostgreSQL Experts Inc.
http://www.pgexperts.com

#3Thom Brown
thom@linux.com
In reply to: Josh Berkus (#2)
Re: Avoiding deadlocks ...

On 20 August 2010 09:39, Josh Berkus <josh@agliodbs.com> wrote:

On 8/19/10 3:51 PM, Josh Berkus wrote:

Kevin,

This one is for you:

Two sessions, in transaction:

Process A             Process B

update session where id = X;
                      update order where orderid = 5;
update order where orderid = 5;
                      update order where orderid = 5;
... deadlock error.

Johto on IRC pointed out I left something out of the above: "session" is
referenced in an FK by "orders", and session = X is related to orderid = 5.

I was wondering what that had to do with anything.

--
Thom Brown
Registered Linux user: #516935

#4Marko Tiikkaja
marko.tiikkaja@cs.helsinki.fi
In reply to: Josh Berkus (#2)
Re: Avoiding deadlocks ...

On 2010-08-20 11:39 AM +0300, Josh Berkus wrote:

On 8/19/10 3:51 PM, Josh Berkus wrote:

Two sessions, in transaction:

Process A Process B

update session where id = X;
update order where orderid = 5;
update order where orderid = 5;
update order where orderid = 5;
... deadlock error.

Johto on IRC pointed out I left something out of the above: "session" is
referenced in an FK by "orders", and session = X is related to orderid = 5.

Right, that would result in a deadlock. I think truly serializable
transactions still need to SELECT FOR SHARE here for foreign keys to
work, no?

Regards,
Marko Tiikkaja

#5Kevin Grittner
Kevin.Grittner@wicourts.gov
In reply to: Marko Tiikkaja (#4)
Re: Avoiding deadlocks ...

Josh Berkus wrote:

Two sessions, in transaction:

Process A Process B

update session where id = X;
update order where orderid = 5;
update order where orderid = 5;
update order where orderid = 5;
... deadlock error.

Johto on IRC pointed out I left something out of the above:
"session" is referenced in an FK by "orders", and session = X is
related to orderid = 5.

The patch I'm offering implements the SSI techniques published by
Michael Cahill, et al. Those techniques basically allow the current
snapshot isolation to run as it currently does, but monitors for
read/write conflicts to generate a new type of serialization failure
when a cycle becomes possible which could create an anomaly. There
are no read/write conflict cycles in your example, so it would behave
just as REPEATABLE READ and SERIALIZABLE now behave -- you get a
deadlock which rolls back one of the transactions.

I don't see how SSI can be modified to generate some other form of
serialization failure here, but I'm always open to suggestions.

-Kevin

#6Kevin Grittner
Kevin.Grittner@wicourts.gov
In reply to: Kevin Grittner (#5)
Re: Avoiding deadlocks ...

I wrote:

I don't see how SSI can be modified to generate some other form of
serialization failure here, but I'm always open to suggestions.

Actually, after thinking about it a bit more, the UPDATE statements
*do* read the rows before writing, so a naive implementation would
see write skew in Josh's example and generate a rollback before
things got far enough to cause a deadlock. In fact, a few months
ago the implementation probably would have done so, before we
implemented the optimization mentioned in section 3.7.3 of Cahill's
doctoral thesis[1]http://hdl.handle.net/2123/5353.

The reasons for implementing that change were:

(1) It avoids getting an SIREAD lock on a row if that row has been
updated by the transaction. I believe that in the PostgreSQL
implementation we even avoid taking the SIREAD lock when we're in a
scan from an UPDATE or DELETE statement, but I'd have to dig into
the code to confirm.

(2) Because of (1) and the removal of an SIREAD lock on a row is
later updated, the shared memory structures used for tracking SIREAD
locks can be somewhat smaller and access to them will be a bit
faster.

(3) I *think* that having the additional SIREAD locks would tend to
increase the false positive rate, although I'd need to spend some
time working through that to be sure.

So, the question would be: does this "optimization" from the paper
actually improve performance because of the above points more than
the savings which would accrue from catching the conflict in Josh's
example before it gets to the point of deadlock? I can add that to
the list of things to check once we have a good set of benchmarks.

-Kevin

[1]: http://hdl.handle.net/2123/5353

#7Kevin Grittner
Kevin.Grittner@wicourts.gov
In reply to: Marko Tiikkaja (#4)
Re: Avoiding deadlocks ...

Marko Tiikkaja <marko.tiikkaja@cs.helsinki.fi> wrote:

I think truly serializable transactions still need to SELECT FOR
SHARE here for foreign keys to work, no?

That depends on how you look at it. The SSI patch that Dan and I
have been working on doesn't attempt to change the implementation
techniques for foreign keys, because SSI only enforces integrity
among serializable transactions -- and we want foreign keys to be
enforced regardless of the transaction isolation levels used.

When writing queries which will be run at the serializable isolation
level, if you are only concerned with anomalies from interaction
with other serializable transactions, you *never* have to explicitly
code SELECT FOR SHARE or SELECT FOR UPDATE, nor do you ever need to
explicitly request a lock; so from that perspective the answer to
the question is "No." Under the covers, PostgreSQL will continue to
use existing techniques for enforcing referential integrity defined
by foreign keys; so from that perspective the answer to the question
is "Yes."

Hopefully that made sense....

-Kevin

#8Marko Tiikkaja
marko.tiikkaja@cs.helsinki.fi
In reply to: Kevin Grittner (#7)
Re: Avoiding deadlocks ...

On 2010-08-20 6:19 PM, Kevin Grittner wrote:

Marko Tiikkaja<marko.tiikkaja@cs.helsinki.fi> wrote:

I think truly serializable transactions still need to SELECT FOR
SHARE here for foreign keys to work, no?

That depends on how you look at it. The SSI patch that Dan and I
have been working on doesn't attempt to change the implementation
techniques for foreign keys, because SSI only enforces integrity
among serializable transactions -- and we want foreign keys to be
enforced regardless of the transaction isolation levels used.

Exactly.

Regards,
Marko Tiikkaja

#9Greg Stark
gsstark@mit.edu
In reply to: Josh Berkus (#1)
Re: Avoiding deadlocks ...

On Thu, Aug 19, 2010 at 11:51 PM, Josh Berkus <josh@agliodbs.com> wrote:

update session where id = X;
                       update order where orderid = 5;
update order where orderid = 5;

So i think this will already deadlock.

A has a exclusive-lock on session<X> and is waiting on order<5>. B has
an exclusive lock on order<5> and is waiting on a share-lock on
session<x>

                       update order where orderid = 5;
... deadlock error.

Do you actually get a prompt here to type this command?

--
greg

#10Kevin Grittner
Kevin.Grittner@wicourts.gov
In reply to: Greg Stark (#9)
Re: Avoiding deadlocks ...

Greg Stark <gsstark@mit.edu> wrote:

Josh Berkus <josh@agliodbs.com> wrote:

update session where id = X;
update order where orderid = 5;
update order where orderid = 5;

So i think this will already deadlock.

A has a exclusive-lock on session<X> and is waiting on order<5>. B
has an exclusive lock on order<5> and is waiting on a share-lock
on session<x>

No, see Tom's explanation on the related "Deadlock bug" thread:

http://archives.postgresql.org/pgsql-hackers/2010-08/msg01464.php

update order where orderid = 5;
... deadlock error.

Do you actually get a prompt here to type this command?

Yes. The attachment at the start of the other thread makes it easy
to confirm:

http://archives.postgresql.org/pgsql-hackers/2010-08/msg01447.php

-Kevin

#11Josh Berkus
josh@agliodbs.com
In reply to: Marko Tiikkaja (#8)
Re: Avoiding deadlocks ...

On 8/20/10 8:23 AM, Marko Tiikkaja wrote:

On 2010-08-20 6:19 PM, Kevin Grittner wrote:

Marko Tiikkaja<marko.tiikkaja@cs.helsinki.fi> wrote:

I think truly serializable transactions still need to SELECT FOR
SHARE here for foreign keys to work, no?

That depends on how you look at it. The SSI patch that Dan and I
have been working on doesn't attempt to change the implementation
techniques for foreign keys, because SSI only enforces integrity
among serializable transactions -- and we want foreign keys to be
enforced regardless of the transaction isolation levels used.

Ok, then that's not a fix for this particular problem. This case is a
good example, though, of showing how deadlocks are the most expensive
type of serialization failure, and thus models which avoid deadlocks (in
favor of other kinds of blocking and/or serialization errors) are desirable.

--
-- Josh Berkus
PostgreSQL Experts Inc.
http://www.pgexperts.com