What is the bitmap index?

示例执行计划：

postgres=# explain select * from tbl where id =1 or id=2 or id=3;
QUERY PLAN
----------------------------------------------------------------------------------
Bitmap Heap Scan on tbl (cost=124.97..354.97 rows=10000 width=12)
Recheck Cond: ((id = 1) OR (id = 2) OR (id = 3))
-> BitmapOr (cost=124.97..124.97 rows=10000 width=0)
-> Bitmap Index Scan on idx_tbl (cost=0.00..114.28 rows=10000 width=0)
Index Cond: (id = 1)
-> Bitmap Index Scan on idx_tbl (cost=0.00..1.59 rows=1 width=0)
Index Cond: (id = 2)
-> Bitmap Index Scan on idx_tbl (cost=0.00..1.59 rows=1 width=0)
Index Cond: (id = 3)
(9 rows) 

> what does "Bitmap Heap Scan" phase do?

A plain indexscan fetches one tuple-pointer at a time from the index,
and immediately visits that tuple in the table. A bitmap scan fetches
all the tuple-pointers from the index in one go, sorts them using an
in-memory "bitmap" data structure, and then visits the table tuples in
physical tuple-location order. The bitmap scan improves locality of
reference to the table at the cost of more bookkeeping overhead to
manage the "bitmap" data structure --- and at the cost that the data
is no longer retrieved in index order, which doesn't matter for your
query but would matter if you said ORDER BY.

> what is "Recheck condition" and why is it needed?

If the bitmap gets too large we convert it to "lossy" style, in which we
only remember which pages contain matching tuples instead of remembering
each tuple individually. When that happens, the table-visiting phase
has to examine each tuple on the page and recheck the scan condition to
see which tuples to return.

> I thought "Bitmap Index Scan" was only used when there are two or more applicable indexes in the plan, so I don't understand why is it used now?

True, we can combine multiple bitmaps via AND/OR operations to merge
results from multiple indexes before visiting the table ... but it's
still potentially worthwhile even for one index. A rule of thumb is
that plain indexscan wins for fetching a small number of tuples, bitmap
scan wins for a somewhat larger number of tuples, and seqscan wins if
you're fetching a large percentage of the whole table.

stackoverflow相关问题

问题

How does PostgreSQL knows by just a bitmap anything about rows' physical order?
回答

The bitmap is one bit per heap page. The bitmap index scan sets the bits based on the heap page address that the index entry points to.

So when it goes to do the bitmap heap scan, it just does a linear table scan, reading the bitmap to see whether it should bother with a particular page or seek over it.
问题

Or generates the bitmap so that any element of it can be mapped to the pointer to a page easily?
回答

No, the bitmap corresponds 1:1 to heap pages.

I wrote some more on this here.

OK, it looks like you might be misunderstanding what "bitmap" means in this context.

It's not a bit string like "101011" that's created for each heap page, or each index read, or whatever.

The whole bitmap is a single bit array, with as many bits as there are heap pages in the relation being scanned.

One bitmap is created by the first index scan, starting off with all entries 0 (false). Whenever an index entry that matches the search condition is found, the heap address pointed to by that index entry is looked up as an offset into the bitmap, and that bit is set to 1 (true). So rather than looking up the heap page directly, the bitmap index scan looks up the corresponding bit position in the bitmap.

The second and further bitmap index scans do the same thing with the other indexes and the search conditions on them.

Then each bitmap is ANDed together. The resulting bitmap has one bit for each heap page, where the bits are true only if they were true in all the individual bitmap index scans, i.e. the search condition matched for every index scan. These are the only heap pages we need to bother to load and examine. Since each heap page might contain multiple rows, we then have to examine each row to see if it matches all the conditions - that's what the "recheck cond" part is about.

One crucial thing to understand with all this is that the tuple address in an index entry points to the row's ctid, which is a combination of the heap page number and the offset within the heap page. A bitmap index scan ignores the offsets, since it'll check the whole page anyway, and sets the bit if any row on that page matches the condition.

Graphical example

Heap, one square = one page:
+---------------------------------------------+
|c____u_____X___u___X_________u___cXcc______u_|
+---------------------------------------------+
Rows marked c match customers pkey condition.
Rows marked u match username condition.
Rows marked X match both conditions.

Bitmap scan from customers_pkey:
+---------------------------------------------+
|100000000001000000010000000000000111100000000| bitmap 1
+---------------------------------------------+
One bit per heap page, in the same order as the heap
Bits 1 when condition matches, 0 if not

Bitmap scan from ix_cust_username:
+---------------------------------------------+
|000001000001000100010000000001000010000000010| bitmap 2
+---------------------------------------------+

Once the bitmaps are created a bitwise AND is performed on them:

+---------------------------------------------+
|100000000001000000010000000000000111100000000| bitmap 1
|000001000001000100010000000001000010000000010| bitmap 2
 &&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&
|000000000001000000010000000000000010000000000| Combined bitmap
+-----------+-------+--------------+----------+
            |       |              |
            v       v              v
Used to scan the heap only for matching pages:
+---------------------------------------------+
|___________X_______X______________X__________|
+---------------------------------------------+

The bitmap heap scan then seeks to the start of each page and reads the page:

+---------------------------------------------+
|___________X_______X______________X__________|
+---------------------------------------------+
seek------->^seek-->^seek--------->^
            |       |              |
            ------------------------
            only these pages read

and each read page is then re-checked against the condition since there can be >1 row per page and not all necessarily match the condition.

---

The bitmap of pages is created dynamically for each query. It is not cached or re-used, and is discarded at the end of the bitmap index scan.

It doesn't make sense to create the page bitmap in advance because its contents depend on the query predicates.

Say you're searching for x=1 and y=2. You have b-tree indexes on x and y. PostgreSQL doesn't combine x and y into a bitmap then search the bitmap. It scans index x for the page address of all pages with x=1 and makes a bitmap where the pages that might contain x=1 are true. Then it scans y looking for the page addresses where y might equal 2, making a bitmap from that. Then it ANDs them to find pages where both x=1 and y=2 might be true. Finally, it scans the table its self, reading only the pages that might contain candidate values, reading each page and keeping only the rows where x=1 and y=2.

Now, if you're looking for something like a cached, pre-built bitmap index, there is such a thing in PostgreSQL 9.5: BRIN indexes. These are intended for very large tables, and provide a way to find ranges of the table that can be skipped over because they're known not to contain a desired value.

---

To combine multiple indexes, the system scans each needed index and prepares a bitmap in memory giving the locations of table rows that are reported as matching that index's conditions. The bitmaps are then ANDed and ORed together as needed by the query. Finally, the actual table rows are visited and returned. The table rows are visited in physical order, because that is how the bitmap is laid out; this means that any ordering of the original indexes is lost, and so a separate sort step will be needed if the query has an ORDER BY clause. For this reason, and because each additional index scan adds extra time, the planner will sometimes choose to use a simple index scan even though additional indexes are available that could have been used as well.

注：

这里关于stackoverflow部分画的图有个疑问，满足where条件的 heap page标为1，不满足的则标为0，然后做的and操作，得出的是同时含有where条件的heap page标记为1，但我查询的目的，并不是要找出记录在同一个heap page中的记录，而是找到满足where条件的记录，只要满足where条件，不必一定要在同一个heap page中的。希望了解的同学告知一下。

参考：

https://www.postgresql.org/message-id/12553.1135634231@sss.pgh.pa.us

https://dba.stackexchange.com/questions/119386/understanding-bitmap-heap-scan-and-bitmap-index-scan

https://stackoverflow.com/questions/33100637/undestanding-bitmap-indexes-in-postgresql

https://yq.aliyun.com/articles/70462

https://www.postgresql.org/docs/9.6/static/indexes-bitmap-scans.html