postgresql locks the whole table!

[Dr NoName]

Help!

I have a table that multiple processes must be able to
write to concurrently. However, it for some reason
gets locked in exclusive mode. I narrowed it down to
one SQL statement + some weirdness with foreign keys.
To debug this, I opened two psql sessions and typed in
the sql statements manually. Here is the situation:

CREATE TABLE take2
(
id serial not null,
timestamp timestamp NOT NULL DEFAULT now(),
description text,
iteration smallint,
asset_id integer,
-- FOREIGN KEY (asset_id) REFERENCES public.asset
(id), -- ON UPDATE CASCADE ON DELETE CASCADE,

primary key(id)
);

(notice that the foreign key statement is commented
out). Just to make sure I am not causing excessive
locking unintentionally , I did "set transaction
isolation level read committed" in both psql shells
(default was serializable).

Now I type the following commands:

shell 1:

1. BEGIN
2. insert into take2 values(default, 'now()', 't1', 1,
1);
shell 2:

1. BEGIN
2. insert into take2 values(default, 'now()', 't2', 1,
1);

this works.

However, if I uncomment the foreign key statement and
recreate the table, then the second shell blocks on
the insert statement. As soon as the first transaction
is either committed or rolled back, the insert
statement goes through.

My question is why??? The two insert operations do not
conflict with each other (at least not in the
real-world situation). Also, why does the foreign key
make a difference?

looking at pg_locks, I see the following:

relation | database | transaction | pid |
mode | granted
----------+----------+-------------+-------+------------------+---------
39356 | 34862 | NULL | 18671 |
AccessShareLock | t
39356 | 34862 | NULL | 18671 |
RowExclusiveLoc k | t
NULL | NULL | 9914 | 18671 |
ExclusiveLock | t
39354 | 34862 | NULL | 18671 |
AccessShareLock | t
34886 | 34862 | NULL | 18671 |
AccessShareLock | t
34886 | 34862 | NULL | 18671 |
RowShareLock | t
16759 | 34862 | NULL | 18671 |
AccessShareLock | t
(7 rows)

Where does the ExclusiveLock come from? What is being
locked?

It is critical for us to run multiple transactions
concurrently -- in fact that was one of the reasons
for choosing PostgreSQL over MySQL. There are a lot of
file system operations and other processing that need
to happen along side the DB transaction. Those things
take a long time, so there is typically up to a
5-minute span between BEGIN and COMMIT. We cannot
block the production floor for 5 minutes when a user
tries to run a transaction, so as a temporary fix, we
got rid of the begin/commit. But obviously we would
rather not lose the atomicity.

So, in summary:
why does PostgreSQL lock the entire table?
what can we do about it?

This was tested on PostgreSQL 7.4.0 and 7.3.2.

thanks in advance,

Eugene

_______________ _______________ ____
Do you Yahoo!?
Free Pop-Up Blocker - Get it now
http://companion.yahoo.com/

---------------------------(end of broadcast)---------------------------
TIP 9: the planner will ignore your desire to choose an index scan if your
joining column's datatypes do not match

[Bruce Momjian]

Greg Stark wrote:

Dr NoName <sp********@yah oo.com> writes:

My question is why??? The two insert operations do not
conflict with each other (at least not in the
real-world situation). Also, why does the foreign key
make a difference?

It's not locking the whole table, it's locking the record that the foreign key
references. Note that they're both referencing the same foreign key.

It does this because it's afraid someone will go and delete that key before
the transaction commits. It has to take a lock that will prevent someone from
deleting the record (or updating the referenced column).

Unfortunately the only lock to choose from is an exclusive write lock. That's
overkill as you've noticed. I think this is something multiple people would
like to fix by introducing shared locks, but I wouldn't expect a solution
soon.

As I remember the implementation problem is that we do an exclusive row
lock right now by putting the transaction id on the row and set a
special row-lock bit in the tuple. Shared locks need to store multiple
transaction ids, and that is where we are stuck. Shared memory is of
finite size, and the number of proc/row combinations is infinite, so it
seems we will need some shared stucture with backing store to disk, and
that will be slow.

You know the maximum number of backends on startup, but I don't see how
that helps us. If we could somehow know the unique combinations of
those backend ids that would be used for any row, we could reserve a
range of transactions ids to map them, but the number of combinations is
too large.

Our xid/command-counter is currently 64 bits, so if we only had a
maximum of 64 backends, we could use those bits to mark the backends
holding the locks rather than put the xid in there. Of course, the
maximum number backends can change over time, so that isn't really a
solution but more a brainstorm, but I do think shared memory bitmaps
might be in the final solution.

One idea would be to allocate 10k of shared memory for a shared lock
bitmap. Assuming a max 100 backend, that is 2500 lock combinations,
numbered 0-2499. We would put the bitmap number on the rows rather than
the xid. Of course, problems are that there are only 2500 combinations
supported, and transactions have to get an exclusive lock on transaction
commit to clear their backend bits from the bitmap table so the rows can
be reused. Another refinement would be to use the row xid to store
either the xid for single backend locks, and use the bitmap table number
only when there is sharing of row locks by multiple backends. That
might reduce the contention on the bitmap table. If a backend wasn't
involved in shared locks, its bit wouldn't be set in the bitmap table
and it has nothing to do, and it can read the table without a lock.

--
Bruce Momjian | http://candle.pha.pa.us
pg***@candle.ph a.pa.us | [phone]
+ If your life is a hard drive, | 13 Roberts Road
+ Christ can be your backup. | Newtown Square, Pennsylvania 19073

---------------------------(end of broadcast)---------------------------
TIP 9: the planner will ignore your desire to choose an index scan if your
joining column's datatypes do not match

[Jeff Davis]

As I remember the implementation problem is that we do an exclusive row
lock right now by putting the transaction id on the row and set a
special row-lock bit in the tuple. Shared locks need to store multiple
transaction ids, and that is where we are stuck. Shared memory is of
finite size, and the number of proc/row combinations is infinite, so it
seems we will need some shared stucture with backing store to disk, and
that will be slow.

You know the maximum number of backends on startup, but I don't see how
that helps us. If we could somehow know the unique combinations of
those backend ids that would be used for any row, we could reserve a
range of transactions ids to map them, but the number of combinations is
too large.

Our xid/command-counter is currently 64 bits, so if we only had a
maximum of 64 backends, we could use those bits to mark the backends
holding the locks rather than put the xid in there. Of course, the
maximum number backends can change over time, so that isn't really a
solution but more a brainstorm, but I do think shared memory bitmaps
might be in the final solution.

One idea would be to allocate 10k of shared memory for a shared lock
bitmap. Assuming a max 100 backend, that is 2500 lock combinations,
numbered 0-2499. We would put the bitmap number on the rows rather than
the xid. Of course, problems are that there are only 2500 combinations
supported, and transactions have to get an exclusive lock on transaction
commit to clear their backend bits from the bitmap table so the rows can
be reused. Another refinement would be to use the row xid to store
either the xid for single backend locks, and use the bitmap table number
only when there is sharing of row locks by multiple backends. That
might reduce the contention on the bitmap table. If a backend wasn't
involved in shared locks, its bit wouldn't be set in the bitmap table
and it has nothing to do, and it can read the table without a lock.

The way I understand it, is that you're having trouble storing all of
the xids in the row; right now you just store one and mark it "locked".
If you were able to store several, but not necessarily all xids in all
cases, wouldn't that be a big improvement? The xids not in the list
would lock, as they do now, and the ones in the list would show an
improvement by sharing the lock, right?

Otherwise I don't entirely understand what you're saying.

Regards,
Jeff Davis

---------------------------(end of broadcast)---------------------------
TIP 7: don't forget to increase your free space map settings

[Bruce Momjian]

Jeff Davis wrote:

The way I understand it, is that you're having trouble storing all of
the xids in the row; right now you just store one and mark it "locked".
If you were able to store several, but not necessarily all xids in all
cases, wouldn't that be a big improvement? The xids not in the list
would lock, as they do now, and the ones in the list would show an
improvement by sharing the lock, right?

Otherwise I don't entirely understand what you're saying.

Yes, we could do that but we really want something that isn't going to
work only some of the time. Ideally we want something that can share no
matter how many backends try.

--
Bruce Momjian | http://candle.pha.pa.us
pg***@candle.ph a.pa.us | [phone]
+ If your life is a hard drive, | 13 Roberts Road
+ Christ can be your backup. | Newtown Square, Pennsylvania 19073

---------------------------(end of broadcast)---------------------------
TIP 8: explain analyze is your friend

[Greg Stark]

It's not strictly necessary to have a list of all xids at all. The normal
"shared read lock" is just "take the write lock, increment the readers
counter, unlock" Anyone who wants to write has to wait (using, eg, a condition
variable) until the readers count goes to 0.

This gets the right semantics but without the debugging info of a list of
lockers. Other than debugging the only advantage I see to having the list of
lockers is for deadlock detection. Is that absolutely mandatory?

--
greg
---------------------------(end of broadcast)---------------------------
TIP 2: you can get off all lists at once with the unregister command
(send "unregister YourEmailAddres sHere" to ma*******@postg resql.org)

[Mike Mascari]

Greg Stark wrote:

It's not strictly necessary to have a list of all xids at all. The normal
"shared read lock" is just "take the write lock, increment the readers
counter, unlock" Anyone who wants to write has to wait (using, eg, a condition
variable) until the readers count goes to 0.

This gets the right semantics but without the debugging info of a list of
lockers. Other than debugging the only advantage I see to having the list of
lockers is for deadlock detection. Is that absolutely mandatory?

What happens if a backend is killed and never decrements its reference
count?

Mike Mascari
ma*****@mascari .com
---------------------------(end of broadcast)---------------------------
TIP 6: Have you searched our list archives?

http://archives.postgresql.org

[Stephan Szabo]

On Sun, 7 Dec 2003, Greg Stark wrote:

It's not strictly necessary to have a list of all xids at all. The normal
"shared read lock" is just "take the write lock, increment the readers
counter, unlock" Anyone who wants to write has to wait (using, eg, a condition
variable) until the readers count goes to 0.
There are some storage/cleanup questions though. If that's stored in the
tuple header, what happens after a crash?

In addition, how should the locks be granted for a sequence like:
T1: get shared lock on row A
T2: get exclusive lock on row A
T3: get shared lock on row A
Does T3 get the lock or not? If it does, then you have the possibility of
freezing out T2 for a very long time and badly hurting update/delete
performance. If it doesn't, then how are you keeping track of the fact
that there are one or more people who want exclusive locks on the same
row that are "in front" of you?
This gets the right semantics but without the debugging info of a list of
lockers. Other than debugging the only advantage I see to having the list of
lockers is for deadlock detection. Is that absolutely mandatory?

I think so, yes, especially if we're going to use it for things like
foreign keys. It's too easy to get into a deadlock with foreign keys
(even when implemented through shared locks) and I think having undetected
deadlocks would be even worse than our current behavior. At least with
the current behavior you get an indication that something is wrong.

---------------------------(end of broadcast)---------------------------
TIP 7: don't forget to increase your free space map settings

[Tom Lane]

> Greg Stark wrote:

This gets the right semantics but without the debugging info of a list of
lockers. Other than debugging the only advantage I see to having the list of
lockers is for deadlock detection. Is that absolutely mandatory?

No, deadlock detection is not optional.

Mike Mascari <ma*****@mascar i.com> writes: What happens if a backend is killed and never decrements its reference
count?

Even if it's not killed, how does it know to decrement the reference
count? You still need a list of all locked tuples *somewhere*. Perhaps
a technique like this would allow the list to not be in shared memory,
which is helpful, but it's far from an ideal solution.

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 8: explain analyze is your friend

[Greg Stark]

Stephan Szabo <ss****@megazon e.bigpanda.com> writes:

In addition, how should the locks be granted for a sequence like:
T1: get shared lock on row A
T2: get exclusive lock on row A
T3: get shared lock on row A
Does T3 get the lock or not? If it does, then you have the possibility of
freezing out T2 for a very long time and badly hurting update/delete
performance.

Well this is a fundamental question that applies to any scheme to handle
shared locks. You get into all sorts of fun stuff like livelock and priority
inversion that real time systems folk invent just to torture programmers.

--
greg
---------------------------(end of broadcast)---------------------------
TIP 2: you can get off all lists at once with the unregister command
(send "unregister YourEmailAddres sHere" to ma*******@postg resql.org)