Details
-
Technical task
-
Status: Closed (View Workflow)
-
Major
-
Resolution: Fixed
-
None
-
None
Description
This task in a part of MDEV-6915, where we would like the pack the values of the sort key inside the sort buffer for each record.
EDIT: this is the spec of what got implemented:
Contents:
1. Background
1.1 Implementation details
1.1.1 Why fields and items are treated differently
2. Solution : Packed Sort Keys
2.1 Packed key format
2.2 Which format to use
3. Special cases
3.1 Handling very long strings
3.2 Handling for long binary strings
3.3 Handling very long strings with Packed sort keys
4. Sort key columns in addon_fields
1. Background
Before this MDEV, filesort() sorted the data using mem-comparable keys.
That is, if we wanted to sort by
ORDER BY col1, col2, ... colN
|
then the filesort code would for each row generate one "Sort Key" and then sort the rows by their Sort Keys.
The Sort Keys are mem-comparable (that is, are compared by memcmp()) and they are fixed size: the sort key has the same length regardless of what value it represents. This causes inefficient memory usage.
1.1 Implementation details
filesort.cc: make_sortkey() is the function that produces a sort key from a record.
The function treats Field and Item objects differently.
class Field has
void make_sort_key(uchar *buff, uint length); |
virtual void sort_string(uchar *buff,uint length)=0; |
sort_string produces mem-comparable image of the field value for each datatype. make_sort_key is a non-virtual function which handles encoding of SQL null values.
For Items, Type_handler has a virtual function:
virtual void make_sort_key(uchar *to, Item *item, |
const SORT_FIELD_ATTR *sort_field, |
Sort_param *param) const= 0; |
which various datatypes overload.
1.1.1 Why fields and items are treated differently
My (Sergey P) guess is as follows: if we use fields we get more compact sort keys. For example:
create table t7(a int); |
select * from t7 order by a; -- Q1 |
select * from t7 order by a+1; -- Q2 |
Q1 uses a Field. It's a Field_int, so the sort key is 4 bytes long.
Q2 uses an Item. Its type handler is Type_handler_int_result, and the sort key is 8 bytes long.
2. Solution : Packed Sort Keys
Note that one can have mem-comparable keys are that are not fixed-size. MyRocks uses such encoding for example.
However for this MDEV it was decided to store the original (non-mem-comparable) values instead, and use a datatype-aware key comparison function. (See the issue comments for the reasoning)
2.1 Packed key format
The keys are stored in a new variable-size data format called "packed".
The format is as follows:
sort_key_length packed_value_1 packed_value2 ...
|
sort_key_length is the length of the whole key.
Each packed value is encoded as follows:
<null_byte=0> // This is a an SQL NULL
|
[<null_byte=1>] <packed_value> // this a non-NULL value
|
null_byte is present if the field/item is NULLable.
SQL NULL is encoded as just one NULL-indicator byte. The value itself is omitted.
The format of the packed_value depends on the datatype. For "non-packable" datatypes it is just their mem-comparable form, as before.
The "packable" datatypes are currently variable-length strings and the packed format for them is (for binary blobs, see a note below):
<length> <string>
|
2.2 Which format to use
The advantage of Packed Key Format is potential space savings for variable-length fields.
The disadvantages are:
- it may actually take more space, because of sort_key_length and length fields.
- The comparison function is more expensive.
Currently the logic is: use Packed Key Format if we would save 20 or more bytes when constructing a sort key from values that have empty string for each packable component.
3. Special cases
3.1 Handling very long strings
Before this MDEV, the size of sort key was limited by @@max_sort_length variable.
It is defined as:
The number of bytes to use when sorting data values. The server uses only the first max_sort_length bytes of each value and ignores the rest.
3.2 Handling for long binary strings
Long binary strings receive special treatment. A sort key for the long binary string is truncated at max_sort_length bytes like described above, but then a "suffix" is appended which contains the total length of the value before the truncation.
3.3 Handling very long strings with Packed sort keys
Truncating multi-byte string at N bytes is not safe because one can cut in the middle of a character.
One is tempted to solve this by discarding the partial character but that's also not a good idea as in some collations multiple characters may produce one weight (this is called "contraction").
This combination of circumstances:
- The string value is very long, so truncation is necessary
- The collation is "complex", so truncation is dangerous
is deemed to be relatively rare so it was decided to just use the non-packed sort keys in this case.
4. Sort key columns in addon_fields
Currently, each sort key column is actually stored twice
1. as part of the sort key
2. in the addon_fields
This made total sense when sort key stored the mem-comparable image (from which one cannot restore the original value in general case). But since we now store the original value, we could also remove it from the addon_fields and further save space.
This is a good idea but it is outside of the scope of this MDEV.
Attachments
Issue Links
- causes
-
MDEV-21946 Server crash in store_length upon GROUP BY WITH ROLLUP with geometry field
- Closed
-
MDEV-31743 Server crash in store_length, assertion failure in Type_handler_string_result::sort_length
- Closed
- includes
-
MDEV-21784 Performance testing for packed sort keys
- Closed
-
MDEV-21791 Packed keys: connect.mysql_index crashes in BB on 32-bit box
- Closed
-
MDEV-21802 Packed keys: assertion failure
- Closed
- is part of
-
MDEV-6915 Allow packed keys and packed values of non-sorted fields in the sort buffer
- Closed
- relates to
-
MDEV-21829 Use packed sort keys in Unique objects
- Stalled
-
MDEV-6181 EITS could eat all tmpdir space and hang
- Closed