On-disk bitmap index implementation

Hi all,

Attached is a patch implementing bitmap indexes. It includes major
enhancements on the patch submitted during feature freeze for 8.2 here[1].

In particular: much better integration with the existing bitmap scan code
with the internals of the bitmap streaming pushed down into the AM and
hidden from the executor code; completely new index creation algorithm
which reduced creation time by 20-75% depending on the data; modifications
to the encoding mechanism to suit the integration with bitmap index scans;
work on memory management; lots of code rewriting; range query support.
The code is also much cleaner now.

There are still some things Jie and I have not gotten to yet:

o Improving VACUUM support -- currently, VACUUM FULL means REINDEX for
bitmaps. Heikki Linnakangas offered to work on this. Heikki, are you
still interested?

o Determine if we need to provide anything for rm_startup, rm_cleanup,
rm_safe_restartpoint RmgrData function pointers.

o Test WAL replay more thoroughly.

o I pulled a nice optimisation out of the bitmap scan OR case where a
higher level plan could push down a bitmap and have all the child scans
just OR their data into that inside the AM. I need to get that back in.

o I need to look at tidying up the bitmap stream memory usage insider the
executor. We leak memory from ExecBitmapIndexReScan(), for example.

o Really should add some more detailed docs about why bitmap indexes are
cool.

o Look into adding an AM option such that the user can determine word size
at index creation time. For higher-cardinality data (above 1000 distinct
values), 16 bit word sizes can really help with performance. Although
the word size is not just assumed to be a certain size across the code,
macros are used extensively to interact with the word size. Making it
different for each index might be a little messy.

Comments please!

Gavin

[1] http://archives.postgresql.org/pgsql-patches/2006-09/msg00216.php

On Tue, 2006-12-05 at 00:18 +1100, Gavin Sherry wrote:

> o Determine if we need to provide anything for rm_startup, rm_cleanup,
> rm_safe_restartpoint RmgrData function pointers.

safe_restartpoint gives true/false based upon whether there are
multi-record WAL states that have only been partially received. For
example, a btree index split needs multiple WAL records as it recurses
up the index tree. If you've got one record but not the others yet you
have an incomplete state and so cannot safely write a restartpoint.

I'll document that if you/anyone might suggest where the best place is?

> o Look into adding an AM option such that the user can determine word size
> at index creation time. For higher-cardinality data (above 1000 distinct
> values), 16 bit word sizes can really help with performance. Although
> the word size is not just assumed to be a certain size across the code,
> macros are used extensively to interact with the word size. Making it
> different for each index might be a little messy.

...and is is it a typical case to have a bitmap with less than 1000
distinct values?? Surely we want that as the sole assumption?

Nearly unique bitmaps can suffer a little I think, if it makes the most
common case faster. But I'd like to see the perf results first, I guess.

--
Simon Riggs
EnterpriseDB http://www.enterprisedb.com

Gavin Sherry wrote:
> Hi all,
>
> Attached is a patch implementing bitmap indexes. It includes major
> enhancements on the patch submitted during feature freeze for 8.2 here[1].
>
> In particular: much better integration with the existing bitmap scan code
> with the internals of the bitmap streaming pushed down into the AM and
> hidden from the executor code; completely new index creation algorithm
> which reduced creation time by 20-75% depending on the data; modifications
> to the encoding mechanism to suit the integration with bitmap index scans;
> work on memory management; lots of code rewriting; range query support.
> The code is also much cleaner now.

Thanks! I'll take a look at it.

We need to give the indexam API some further thought. As you know, I've
been working on the Grouped Index Tuples stuff, which also requires
changes to the API to get full benefit. There's a bunch of functionality
I'd like to see:

* Support for streamed bitmaps, like you have implemented.

* Support for candidate matches. This is needed for GIT, as well as
range-encoded bitmap indexes if/when we add them.

* Support for returning tuples in partial order. This is again needed
for GIT, because grouped tuples don't keep track of the ordering within
the group, so they need to be sorted if ordering necessary. And again
it's also useful to return tuples in order from range-encoded bitmaps.

* Support for kill_prior_tuple on bitmap scans.

* A bulk insert API. When inserting a lot of tuples with similar keys,
we could a considerable amount of CPU with a bulk insert API. A bulk
insert to a B-tree for example would only need to descend the tree once,
find the insert location, lock the target page just once and insert all
the tuples that belong to that page. That would potentially also reduce
WAL traffic.

> There are still some things Jie and I have not gotten to yet:
>
> o Improving VACUUM support -- currently, VACUUM FULL means REINDEX for
> bitmaps. Heikki Linnakangas offered to work on this. Heikki, are you
> still interested?

Yeah, I can look into that.

> o Test WAL replay more thoroughly.

I've had that problem too with a lot of things I've hacked. I've used a
shell script that does the operation under test, runs a select, kills
and restarts postmaster, and reruns the select. If the select after
crash returns the same result as before, presumably WAL code works. But
you need to watch out for full page writes that might mask bugs in the
redo code.

Anyone have a more sophisticated method?

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com

On 12/4/06 8:22 AM, "Heikki Linnakangas" <heikki(at)enterprisedb(dot)com> wrote:

> Gavin Sherry wrote:
>> Hi all,
>>
>> Attached is a patch implementing bitmap indexes. It includes major
>> enhancements on the patch submitted during feature freeze for 8.2 here[1].
>>
>> In particular: much better integration with the existing bitmap scan code
>> with the internals of the bitmap streaming pushed down into the AM and
>> hidden from the executor code; completely new index creation algorithm
>> which reduced creation time by 20-75% depending on the data; modifications
>> to the encoding mechanism to suit the integration with bitmap index scans;
>> work on memory management; lots of code rewriting; range query support.
>> The code is also much cleaner now.
>
> Thanks! I'll take a look at it.

Thanks!

>
> We need to give the indexam API some further thought. As you know, I've
> been working on the Grouped Index Tuples stuff, which also requires
> changes to the API to get full benefit. There's a bunch of functionality
> I'd like to see:
>
> * Support for streamed bitmaps, like you have implemented.

Yes. It is also interesting to see how this type of stream bitmaps works the
one from the bitmap index.

>
> * Support for candidate matches. This is needed for GIT, as well as
> range-encoded bitmap indexes if/when we add them.
>

We have replace amgetmulti with amgetbitmap. This should be supported by
setting PagetableEntry accordingly.

> * Support for returning tuples in partial order. This is again needed
> for GIT, because grouped tuples don't keep track of the ordering within
> the group, so they need to be sorted if ordering necessary. And again
> it's also useful to return tuples in order from range-encoded bitmaps.

The bitmap index does not guarantee that returning tuples are ordered, but
it guarantees that the tuples from the same heap page will be returned
consecutively. I don't quite understand why this is useful for
range-encoding bitmaps in particular.

>
> * Support for kill_prior_tuple on bitmap scans.
>
> * A bulk insert API. When inserting a lot of tuples with similar keys,
> we could a considerable amount of CPU with a bulk insert API. A bulk
> insert to a B-tree for example would only need to descend the tree once,
> find the insert location, lock the target page just once and insert all
> the tuples that belong to that page. That would potentially also reduce
> WAL traffic.
>

Yes. Currently during the bitmap index creation, we maintain a buffer to
buffer many inserting tuples before writing them to bitmap pages, which
improves the creation performance by 30%-200% depending on the cardinalities
of the attribute to be indexed. This bulk insert API can take advantage of
this as well.

Thanks,
Jie

Heikki Linnakangas wrote:
> We need to give the indexam API some further thought. As you know, I've
> been working on the Grouped Index Tuples stuff, which also requires
> changes to the API to get full benefit. There's a bunch of functionality
> I'd like to see:
>
> * Support for streamed bitmaps, like you have implemented.
>
> * Support for candidate matches. This is needed for GIT, as well as
> range-encoded bitmap indexes if/when we add them.
>
> * Support for returning tuples in partial order. This is again needed
> for GIT, because grouped tuples don't keep track of the ordering within
> the group, so they need to be sorted if ordering necessary. And again
> it's also useful to return tuples in order from range-encoded bitmaps.
>
> * Support for kill_prior_tuple on bitmap scans.
>
> * A bulk insert API. When inserting a lot of tuples with similar keys,
> we could a considerable amount of CPU with a bulk insert API. A bulk
> insert to a B-tree for example would only need to descend the tree once,
> find the insert location, lock the target page just once and insert all
> the tuples that belong to that page. That would potentially also reduce
> WAL traffic.

Forgot one:

* Ability return index tuple contents, not just pointers to heap, to
allow the executor to use the values stored in the index, see
http://archives.postgresql.org/pgsql-performance/2006-09/msg00080.php

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com

On Mon, 4 Dec 2006, Simon Riggs wrote:

> On Tue, 2006-12-05 at 00:18 +1100, Gavin Sherry wrote:
>
> > o Determine if we need to provide anything for rm_startup, rm_cleanup,
> > rm_safe_restartpoint RmgrData function pointers.
>
> safe_restartpoint gives true/false based upon whether there are
> multi-record WAL states that have only been partially received. For
> example, a btree index split needs multiple WAL records as it recurses
> up the index tree. If you've got one record but not the others yet you
> have an incomplete state and so cannot safely write a restartpoint.
>
> I'll document that if you/anyone might suggest where the best place is?

transam/README ?

>
> > o Look into adding an AM option such that the user can determine word size
> > at index creation time. For higher-cardinality data (above 1000 distinct
> > values), 16 bit word sizes can really help with performance. Although
> > the word size is not just assumed to be a certain size across the code,
> > macros are used extensively to interact with the word size. Making it
> > different for each index might be a little messy.
>
> ...and is is it a typical case to have a bitmap with less than 1000
> distinct values?? Surely we want that as the sole assumption?
>
> Nearly unique bitmaps can suffer a little I think, if it makes the most
> common case faster. But I'd like to see the perf results first, I guess.

I'll put together some performance data on different word sizes.

Thanks,

Gavin

On Mon, 4 Dec 2006, Heikki Linnakangas wrote:

> > o Test WAL replay more thoroughly.
>
> I've had that problem too with a lot of things I've hacked. I've used a
> shell script that does the operation under test, runs a select, kills
> and restarts postmaster, and reruns the select. If the select after
> crash returns the same result as before, presumably WAL code works. But
> you need to watch out for full page writes that might mask bugs in the
> redo code.
>
> Anyone have a more sophisticated method?

Well, I've done a combination of what you did and replaying a bunch of
operations using PITR.

Thanks,

Gavin

Gavin Sherry wrote:
> o Improving VACUUM support -- currently, VACUUM FULL means REINDEX for
> bitmaps. Heikki Linnakangas offered to work on this. Heikki, are you
> still interested?

BTW vacuuming seems quite broken as it is:

hlinnaka(at)heikkilaptop:~/pgsql.bitmap$ ~/pgsql.bitmap/bin/psql -a
postgres < vacuumtest.sql
drop table if exists test;
DROP TABLE
create table test (key int);
CREATE TABLE
create index test_bm on test using bitmap (key);
CREATE INDEX
insert into test values (1);
INSERT 0 1
delete from test;
DELETE 1
vacuum test;
VACUUM
insert into test values (2);
INSERT 0 1
select * from test where key = 1;
key
-----
2
(1 row)

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com

On Tue, 5 Dec 2006, Heikki Linnakangas wrote:

> Gavin Sherry wrote:
> > o Improving VACUUM support -- currently, VACUUM FULL means REINDEX for
> > bitmaps. Heikki Linnakangas offered to work on this. Heikki, are you
> > still interested?
>
> BTW vacuuming seems quite broken as it is:
>
> hlinnaka(at)heikkilaptop:~/pgsql.bitmap$ ~/pgsql.bitmap/bin/psql -a
> postgres < vacuumtest.sql
> drop table if exists test;
> DROP TABLE
> create table test (key int);
> CREATE TABLE
> create index test_bm on test using bitmap (key);
> CREATE INDEX
> insert into test values (1);
> INSERT 0 1
> delete from test;
> DELETE 1
> vacuum test;
> VACUUM
> insert into test values (2);
> INSERT 0 1
> select * from test where key = 1;
> key
> -----
> 2
> (1 row)

Oops :-).

Thanks for pointing it out. I think I might have busted something merging
with HEAD. Don't you hate that?

Thanks,

Gavin

Gavin Sherry wrote:
> There are still some things Jie and I have not gotten to yet:
> ...
> o Test WAL replay more thoroughly.

Found one WAL related bug:

postgres=# CREATE TABLE test (i int);
CREATE TABLE
postgres=# INSERT INTO test SELECT a FROM generate_series(1,10) a;
INSERT 0 10
postgres=# CREATE INDEX mdx ON test USING bitmap(i);
CREATE INDEX
postgres=# INSERT INTO test VALUES (11);
INSERT 0 1
postgres=# \q

killall -9 postgres, and restart. Redo fails with:

PANIC: bm_insert_redo: LOV item is not inserted in pos 2(requested 12)
CONTEXT: xlog redo insert a new LOV item: rel 1663/10817/16388

I haven't dug deeper yet.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com

On Wed, 27 Dec 2006, Heikki Linnakangas wrote:

> Gavin Sherry wrote:
> > There are still some things Jie and I have not gotten to yet:
> > ...
> > o Test WAL replay more thoroughly.
>
> Found one WAL related bug:
>
> postgres=# CREATE TABLE test (i int);
> CREATE TABLE
> postgres=# INSERT INTO test SELECT a FROM generate_series(1,10) a;
> INSERT 0 10
> postgres=# CREATE INDEX mdx ON test USING bitmap(i);
> CREATE INDEX
> postgres=# INSERT INTO test VALUES (11);
> INSERT 0 1
> postgres=# \q
>
> killall -9 postgres, and restart. Redo fails with:
>
> PANIC: bm_insert_redo: LOV item is not inserted in pos 2(requested 12)
> CONTEXT: xlog redo insert a new LOV item: rel 1663/10817/16388
>
> I haven't dug deeper yet.

Yes, there were a bunch of WAL issues we wanted to address. Jie has been
working on this too. Thanks for the feedback, we can use this as a test.

Thanks,

Gavin