Allow ILIKE forward matching to use btree index

On Fri, 20 Dec 2024 03:22:26 +0900
Yugo Nagata <nagata@sraoss.co.jp> wrote:

Hi,

Currently, btree indexes cannot used for ILIKE (~~*) operator if the pattern
has case-varying characters although LIKE (~~) expression can be converted
to indexable clauses by the planner support function (if the collation
is "C" or operator class 'text_pattern_ops' is used).

For example, "t ~~ '123foo%'" is converted to "(t >= '123foo' AND t < '123fop')"
and index scan can be used for this condition. On the other hand, "t ~~* '123foo'"
cannot be converted and sequential scan is used.

Even in this case, we can use a bitmap index scan for the condition
"(t >= '123f' AND t < '123g') OR "(t >= '123F' AND t < '123G')" followed by
recheck by the original condition "t ~~* '123foo'", and this could be faster
than seqscan.

However, even if the support function would return OR clauses, the current
planner implementation cannot not build bitmap scan paths using them.

The attached patches allow to ILIKE (~~*) forward matching to use btree index.

The patch 0001 enables get_index_clause_from_support to receive OR clauses
from support functions and use them to build bitmap index scan later. OR clauses
returned by support functions are collected in the new argument 'orclauses" of
match_restriction_clauses_to_index(), and they are passed to
generate_bitmap_or_paths() later to build bitmap scan paths.

The patch 0002 modifies the support function to return OR clauses as an example
above when ILIKE's pattern has case-varying characters in forward matching. The
OR clauses contains two index conditions for the upper and the lower letter of
the first case-varying character in the pattern respectively. Although the
subsequent characters are not considered in the index scans, it could enough be
faster then sequent scan.

Here is an example.

1. Create a table with random text records

=# CREATE TABLE tbl (t text);
=# INSERT INTO tbl SELECT CASE WHEN i%2=1 THEN upper(x) ELSE x END
FROM (SELECT i, md5(i::text) x FROM generate_series(1,5000000) i);

2. Create an index
=# CREATE INDEX ON tbl (t text_pattern_ops);

3. Before applying patch: Seq Scan was selected. It takes about 4 sec.

=# EXPLIN ANALYZE SELECT * FROM tbl WHERE t ILIKE '12ab%';

I am sorry, the example in my previous main was wrong. I showed the plan
with enable_index_scan = off. Indead, if the pattern starts with
case-insensitive characters like '12', an index scan can be used even with
ILIKE.

postgres=# EXPLAIN ANALYZE SELECT * FROM tbl WHERE t ILIKE 'abc%';
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------
Gather (cost=1000.00..69129.61 rows=500 width=33) (actual time=36.317..4034.770 rows=1188 loops=1)
Workers Planned: 2
Workers Launched: 2
Buffers: shared hit=99 read=41939
-> Parallel Seq Scan on tbl (cost=0.00..68079.61 rows=208 width=33) (actual time=19.908..4023.668 rows=396 loops=3)
Filter: (t ~~* 'abc%'::text)
Rows Removed by Filter: 1666271
Buffers: shared hit=99 read=41939
Planning Time: 0.726 ms
Execution Time: 4035.101 ms
(10 rows)

4. After applying patch: Bitmap Index Scan was selected.

postgres=# EXPLAIN ANALYZE SELECT * FROM tbl WHERE t ILIKE 'abc%';
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------------
Bitmap Heap Scan on tbl (cost=12563.66..58314.53 rows=500 width=33) (actual time=156.485..1266.789 rows=1188 loops=1)
Recheck Cond: (((t ~>=~ 'A'::text) AND (t ~<~ 'B'::text) AND (t ~~* 'abc%'::text)) OR ((t ~>=~ 'a'::text) AND (t ~<~ 'b'::text) AND (t ~~* 'abc%'::text)))
Filter: (t ~~* 'abc%'::text)
Rows Removed by Filter: 311473
Heap Blocks: exact=42010
Buffers: shared hit=1 read=44600
-> BitmapOr (cost=12563.66..12563.66 rows=297029 width=0) (actual time=136.979..136.980 rows=0 loops=1)
Buffers: shared hit=1 read=2590
-> Bitmap Index Scan on tbl_t_idx (cost=0.00..6281.71 rows=148515 width=0) (actual time=80.548..80.549 rows=158375 loops=1)
Index Cond: ((t ~>=~ 'A'::text) AND (t ~<~ 'B'::text))
Buffers: shared read=1272
-> Bitmap Index Scan on tbl_t_idx (cost=0.00..6281.71 rows=148515 width=0) (actual time=56.427..56.427 rows=157042 loops=1)
Index Cond: ((t ~>=~ 'a'::text) AND (t ~<~ 'b'::text))
Buffers: shared hit=1 read=1318
Planning Time: 0.592 ms
Execution Time: 1267.162 ms
(16 rows)

What do you think about it?

I think another application of OR-clause returning support function might be
allowing to use an index for NOT LIKE (!~~) expression because, for example,
"t !~~ '123foo%'" could be converted to "(t < '123foo' OR t >= '123fooz')".
(The second condition is a bit loose but this would be safe and not problem
since tuples are filtered by the original condition after the index scan.)
However, it would not very useful unless the distribution is much skew so that
NOT LIKE expression's selectivity is enough small.

Regards,
Yugo Nagata

--
Yugo Nagata <nagata@sraoss.co.jp>

On Tue, 24 Dec 2024 16:04:42 +0900
Yugo Nagata <nagata@sraoss.co.jp> wrote:

On Fri, 20 Dec 2024 03:22:26 +0900
Yugo Nagata <nagata@sraoss.co.jp> wrote:

Hi,

Currently, btree indexes cannot used for ILIKE (~~*) operator if the pattern
has case-varying characters although LIKE (~~) expression can be converted
to indexable clauses by the planner support function (if the collation
is "C" or operator class 'text_pattern_ops' is used).

For example, "t ~~ '123foo%'" is converted to "(t >= '123foo' AND t < '123fop')"
and index scan can be used for this condition. On the other hand, "t ~~* '123foo'"
cannot be converted and sequential scan is used.

Even in this case, we can use a bitmap index scan for the condition
"(t >= '123f' AND t < '123g') OR "(t >= '123F' AND t < '123G')" followed by
recheck by the original condition "t ~~* '123foo'", and this could be faster
than seqscan.

However, even if the support function would return OR clauses, the current
planner implementation cannot not build bitmap scan paths using them.

The attached patches allow to ILIKE (~~*) forward matching to use btree index.

The patch 0001 enables get_index_clause_from_support to receive OR clauses
from support functions and use them to build bitmap index scan later. OR clauses
returned by support functions are collected in the new argument 'orclauses" of
match_restriction_clauses_to_index(), and they are passed to
generate_bitmap_or_paths() later to build bitmap scan paths.

The patch 0002 modifies the support function to return OR clauses as an example
above when ILIKE's pattern has case-varying characters in forward matching. The
OR clauses contains two index conditions for the upper and the lower letter of
the first case-varying character in the pattern respectively. Although the
subsequent characters are not considered in the index scans, it could enough be
faster then sequent scan.

Here is an example.

1. Create a table with random text records

=# CREATE TABLE tbl (t text);
=# INSERT INTO tbl SELECT CASE WHEN i%2=1 THEN upper(x) ELSE x END
FROM (SELECT i, md5(i::text) x FROM generate_series(1,5000000) i);

2. Create an index
=# CREATE INDEX ON tbl (t text_pattern_ops);

3. Before applying patch: Seq Scan was selected. It takes about 4 sec.

=# EXPLIN ANALYZE SELECT * FROM tbl WHERE t ILIKE '12ab%';

I am sorry, the example in my previous main was wrong. I showed the plan
with enable_index_scan = off. Indead, if the pattern starts with
case-insensitive characters like '12', an index scan can be used even with
ILIKE.

postgres=# EXPLAIN ANALYZE SELECT * FROM tbl WHERE t ILIKE 'abc%';
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------
Gather (cost=1000.00..69129.61 rows=500 width=33) (actual time=36.317..4034.770 rows=1188 loops=1)
Workers Planned: 2
Workers Launched: 2
Buffers: shared hit=99 read=41939
-> Parallel Seq Scan on tbl (cost=0.00..68079.61 rows=208 width=33) (actual time=19.908..4023.668 rows=396 loops=3)
Filter: (t ~~* 'abc%'::text)
Rows Removed by Filter: 1666271
Buffers: shared hit=99 read=41939
Planning Time: 0.726 ms
Execution Time: 4035.101 ms
(10 rows)

4. After applying patch: Bitmap Index Scan was selected.

postgres=# EXPLAIN ANALYZE SELECT * FROM tbl WHERE t ILIKE 'abc%';
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------------
Bitmap Heap Scan on tbl (cost=12563.66..58314.53 rows=500 width=33) (actual time=156.485..1266.789 rows=1188 loops=1)
Recheck Cond: (((t ~>=~ 'A'::text) AND (t ~<~ 'B'::text) AND (t ~~* 'abc%'::text)) OR ((t ~>=~ 'a'::text) AND (t ~<~ 'b'::text) AND (t ~~* 'abc%'::text)))
Filter: (t ~~* 'abc%'::text)
Rows Removed by Filter: 311473
Heap Blocks: exact=42010
Buffers: shared hit=1 read=44600
-> BitmapOr (cost=12563.66..12563.66 rows=297029 width=0) (actual time=136.979..136.980 rows=0 loops=1)
Buffers: shared hit=1 read=2590
-> Bitmap Index Scan on tbl_t_idx (cost=0.00..6281.71 rows=148515 width=0) (actual time=80.548..80.549 rows=158375 loops=1)
Index Cond: ((t ~>=~ 'A'::text) AND (t ~<~ 'B'::text))
Buffers: shared read=1272
-> Bitmap Index Scan on tbl_t_idx (cost=0.00..6281.71 rows=148515 width=0) (actual time=56.427..56.427 rows=157042 loops=1)
Index Cond: ((t ~>=~ 'a'::text) AND (t ~<~ 'b'::text))
Buffers: shared hit=1 read=1318
Planning Time: 0.592 ms
Execution Time: 1267.162 ms
(16 rows)

What do you think about it?

I think another application of OR-clause returning support function might be
allowing to use an index for NOT LIKE (!~~) expression because, for example,
"t !~~ '123foo%'" could be converted to "(t < '123foo' OR t >= '123fooz')".
(The second condition is a bit loose but this would be safe and not problem
since tuples are filtered by the original condition after the index scan.)
However, it would not very useful unless the distribution is much skew so that
NOT LIKE expression's selectivity is enough small.

I added a new patch 0003 that enables ILIKE forward matching to use a SP-Gist index.
Similar to 0002, this generates BitmapOr Index Scan for two index clauses that use
"^@" operator for upper letter and lower letter pattern respectively.

* Before applying the patch:

postgres=# explain analyze select* from tbl where t ilike 'abc%';
QUERY PLAN
----------------------------------------------------------------------------------------------------------------------
Gather (cost=1000.00..18899.52 rows=101 width=33) (actual time=41.799..930.382 rows=253 loops=1)
Workers Planned: 2
Workers Launched: 2
Buffers: shared hit=96 read=12533
-> Parallel Seq Scan on tbl (cost=0.00..17889.42 rows=42 width=33) (actual time=26.041..917.570 rows=84 loops=3)
Filter: (t ~~* 'abc%'::text)
Rows Removed by Filter: 336582
Buffers: shared hit=96 read=12533
Planning Time: 0.690 ms
Execution Time: 930.545 ms
(10 rows)

* After applying the patch:

postgres=# explain analyze select* from tbl where t ilike 'abc%';
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------
Bitmap Heap Scan on tbl (cost=3307.96..16702.11 rows=101 width=33) (actual time=69.449..215.636 rows=253 loops=1)
Recheck Cond: (((t ^@ 'A'::text) AND (t ~~* 'abc%'::text)) OR ((t ^@ 'a'::text) AND (t ~~* 'abc%'::text)))
Filter: (t ~~* 'abc%'::text)
Rows Removed by Filter: 62992
Heap Blocks: exact=12437
Buffers: shared hit=18529
-> BitmapOr (cost=3307.96..3307.96 rows=61212 width=0) (actual time=62.567..62.568 rows=0 loops=1)
Buffers: shared hit=6092
-> Bitmap Index Scan on tbl_t_idx (cost=0.00..1653.96 rows=30606 width=0) (actual time=41.893..41.893 rows=31536 loops=1)
Index Cond: (t ^@ 'A'::text)
Buffers: shared hit=2461
-> Bitmap Index Scan on tbl_t_idx (cost=0.00..1653.96 rows=30606 width=0) (actual time=20.671..20.671 rows=31709 loops=1)
Index Cond: (t ^@ 'a'::text)
Buffers: shared hit=3631
Planning Time: 1.414 ms
Execution Time: 215.903 ms
(16 rows)

--
Yugo NAGATA <nagata@sraoss.co.jp>

On Wed, 2025-01-15 at 01:40 +0900, Yugo NAGATA wrote:

For example, "t ~~ '123foo%'" is converted to "(t >= '123foo' AND
t < '123fop')"
and index scan can be used for this condition. On the other hand,
"t ~~* '123foo'"
cannot be converted and sequential scan is used.

Even in this case, we can use a bitmap index scan for the
condition
"(t >= '123f' AND t < '123g') OR "(t >= '123F' AND t < '123G')"
followed by
recheck by the original condition "t ~~* '123foo'", and this
could be faster
than seqscan.

In theory, there could be many OR clauses:

(t >= '123foo' AND t < '123fop') OR
(t >= '123foo' AND t < '123fop') OR
(t >= '123foo' AND t < '123fop') OR
(t >= '123foo' AND t < '123fop') OR
(t >= '123foo' AND t < '123fop') OR
(t >= '123foo' AND t < '123fop') OR

The patch 0001 enables get_index_clause_from_support to receive
OR clauses
from support functions and use them to build bitmap index scan
later. OR clauses
returned by support functions are collected in the new argument
'orclauses" of
match_restriction_clauses_to_index(), and they are passed to
generate_bitmap_or_paths() later to build bitmap scan paths.

If I understand correctly, this

My apologies, I sent the previous email prematurely. Let me try again:

On Wed, 2025-01-15 at 14:34 -0800, Jeff Davis wrote:

On Wed, 2025-01-15 at 01:40 +0900, Yugo NAGATA wrote:

For example, "t ~~ '123foo%'" is converted to "(t >= '123foo'
AND
t < '123fop')"
and index scan can be used for this condition. On the other
hand,
"t ~~* '123foo'"
cannot be converted and sequential scan is used.

Even in this case, we can use a bitmap index scan for the
condition
"(t >= '123f' AND t < '123g') OR "(t >= '123F' AND t < '123G')"
followed by
recheck by the original condition "t ~~* '123foo'", and this
could be faster
than seqscan.

In theory, there could be many OR clauses:

(t >= '123foo' AND t < '123fop') OR
(t >= '123Foo' AND t < '123Fop') OR
(t >= '123fOo' AND t < '123fOp') OR
(t >= '123FOo' AND t < '123FOp') OR
...

How should that be limited?

Regards,
Jeff Davis

On Wed, 15 Jan 2025 14:40:19 -0800
Jeff Davis <pgsql@j-davis.com> wrote:

My apologies, I sent the previous email prematurely. Let me try again:

On Wed, 2025-01-15 at 14:34 -0800, Jeff Davis wrote:

On Wed, 2025-01-15 at 01:40 +0900, Yugo NAGATA wrote:

For example, "t ~~ '123foo%'" is converted to "(t >= '123foo'
AND
t < '123fop')"
and index scan can be used for this condition. On the other
hand,
"t ~~* '123foo'"
cannot be converted and sequential scan is used.

Even in this case, we can use a bitmap index scan for the
condition
"(t >= '123f' AND t < '123g') OR "(t >= '123F' AND t < '123G')"
followed by
recheck by the original condition "t ~~* '123foo'", and this
could be faster
than seqscan.

In theory, there could be many OR clauses:

(t >= '123foo' AND t < '123fop') OR
(t >= '123Foo' AND t < '123Fop') OR
(t >= '123fOo' AND t < '123fOp') OR
(t >= '123FOo' AND t < '123FOp') OR
...

How should that be limited?

Instead of generating complete patterns considering every case-varying characters,
two clauses considering only the first case-varying character are generated.

For example, for the condition "t ILIKE '123foo%'", the generated condition is
"(t >= '123f' AND t < '123g') OR "(t >= '123F' AND t < '123G')".

Rows meeting "(t >= '123f' AND t < '123g')" includes those whose "t" start with
'123f', that is meeting the following;

(t >= '123foo' AND t < '123fop') OR
(t >= '123fOo' AND t < '123fOp') OR
...

, and rows meeting "(t >= '123F' AND t < '123G')" includes those whose "t" start
with '123F', that is meeting the following;

(t >= '123Foo' AND t < '123Fop') OR
(t >= '123FOo' AND t < '123FOp') OR
...

It is required that the second case-varying character and later are checked after
the index scan, and some rows are filtered, but it will be still faster than full
sequential scan.

Regards,
Yugo Nagata

--
Yugo NAGATA <nagata@sraoss.co.jp>

On Thu, 2025-01-16 at 14:53 +0900, Yugo NAGATA wrote:

Instead of generating complete patterns considering every case-
varying characters,
two clauses considering only the first case-varying character are
generated.

Did you consider other approaches that integrate more deeply into the
indexing infrastructure? This feels almost like a skip scan, which
Petter Geoghegan is working on. If you model the disjunctions as skips,
and provide the right API that the index AM can use, then there would
be no limit.

For example:

start scanning at '123FOO'
when you encounter '123FOP' skip to '123FOo'
continue scanning
when you encounter '123FOp' skip to '123FoO'
...

Also, I'm working on better Unicode support to handle multiple case
variants in patterns. For instance, the Greek letter Sigma has three
case variants (one upper and two lower). What's the right API so that
the index AM knows which case variants will sort first, so that the
skips don't go backward?

Regards,
Jeff Davis

On Thu, 16 Jan 2025 09:41:35 -0800
Jeff Davis <pgsql@j-davis.com> wrote:

On Thu, 2025-01-16 at 14:53 +0900, Yugo NAGATA wrote:

Instead of generating complete patterns considering every case-
varying characters,
two clauses considering only the first case-varying character are
generated.

Did you consider other approaches that integrate more deeply into the
indexing infrastructure? This feels almost like a skip scan, which
Petter Geoghegan is working on. If you model the disjunctions as skips,
and provide the right API that the index AM can use, then there would
be no limit.

For example:

start scanning at '123FOO'
when you encounter '123FOP' skip to '123FOo'
continue scanning
when you encounter '123FOp' skip to '123FoO'
...

That seems true there is similarity between ILIKE search and skip scan,
although, on my understand, skip scan is for multi-column indexes
rather than text. I have no concrete idea how the infrastructure for skip
scan to improve ILIKE, anyway.

Also, I'm working on better Unicode support to handle multiple case
variants in patterns. For instance, the Greek letter Sigma has three
case variants (one upper and two lower). What's the right API so that
the index AM knows which case variants will sort first, so that the
skips don't go backward?

That seems true there is similarity between ILIKE search and skip scan,
although, on my understand, skip scan is for multi-column indexes
rather than text. I have no concrete idea how the infrastructure for skip
scan to improve ILIKE, anyway.

Also, I'm working on better Unicode support to handle multiple case
variants in patterns. For instance, the Greek letter Sigma has three
case variants (one upper and two lower). What's the right API so that
the index AM knows which case variants will sort first, so that the
skips don't go backward?

I don't have idea for handling letters with multiple case variants in my patch,
either. I overlooked such pattern at all. So, I'll withdraw the proposal, for now.

Regards,
Yugo Nagata

--
Yugo NAGATA <nagata@sraoss.co.jp>