use int16_t instead of floats, they're faster and smaller.
but perform intermediate SIMD calculations with floats to avoid overflows.
recall drop with such scheme is below 0.002, often none.
int8_t would've been better but the precision loss is too big
and recall degrades too much.
When the source row is deleted, mark the corresponding node in HNSW
index by setting `tref` to null. An index is added for the `tref` in
secondary table for faster searching of the to-be-marked nodes.
The nodes marked as deleted will still be used for search, but will not
be included in the final query results.
As skipping deleted nodes and not adding deleted nodes for new-inserted
nodes' neighbor list could impact the performance, we now only skip
these nodes in search results.
- for some reason the bitmap is not set for hlindex during the delete so
I had to temporarily comment out one line
All new code of the whole pull request, including one or several files
that are either new files or modified ones, are contributed under the
BSD-new license. I am contributing on behalf of my employer Amazon Web
Services, Inc.
* preserve the graph in memory between statements
* keep it in a TABLE_SHARE, available for concurrent searches
* nodes are generally read-only, walking the graph doesn't change them
* distance to target is cached, calculated only once
* SIMD-optimized bloom filter detects visited nodes
* nodes are stored in an array, not List, to better utilize bloom filter
* auto-adjusting heuristic to estimate the number of visited nodes
(to configure the bloom filter)
* many threads can concurrently walk the graph. MEM_ROOT and Hash_set
are protected with a mutex, but walking doesn't need them
* up to 8 threads can concurrently load nodes into the cache,
nodes are partitioned into 8 mutexes (8 is chosen arbitrarily, might
need tuning)
* concurrent editing is not supported though
* this is fine for MyISAM, TL_WRITE protects the TABLE_SHARE and the
graph (note that TL_WRITE_CONCURRENT_INSERT is not allowed, because an
INSERT into the main table means multiple UPDATEs in the graph)
* InnoDB uses secondary transaction-level caches linked in a list in
in thd->ha_data via a fake handlerton
* on rollback the secondary cache is discarded, on commit nodes
from the secondary cache are invalidated in the shared cache
while it is exclusively locked
* on savepoint rollback both caches are flushed. this can be improved
in the future with a row visibility callback
* graph size is controlled by @@mhnsw_cache_size, the cache is flushed
when it reaches the threshold
instead of one row per node per layer, have one row per node.
store all neighbors for all layers in that row, and the vector itself too
it completely avoids searches in the graph table and
will allow to implement deletions in the future
1. introduce alpha. the value of 1.1 is optimal, so hard-code it.
2. hard-code ef_construction=10, best by test
3. rename hnsw_max_connection_per_layer to mhnsw_max_edges_per_node
(max_connection is rather ambiguous in MariaDB) and add a help text
4. rename hnsw_ef_search to mhnsw_min_limit and add a help text
* mhnsw:
* use primary key, innodb loves and (and the index cannot have dupes anyway)
* MyISAM is ok with that, performance-wise
* must be ha_rnd_init(0) because we aren't going to scan
* MyISAM resets the position on ha_rnd_init(0) so query it before
* oh, and use the correct handler, just in case
* HA_ERR_RECORD_IS_THE_SAME is no error
* innodb:
* return ref_length on create
* don't assume table->pos_in_table_list is set
* ok, assume away, but only for system versioned tables
* set alter_info on create (InnoDB needs to check for FKs)
* pair external_lock/external_unlock correctly
Now there's an FVector class which is a pure vector, an array of floats.
It doesn't necessarily corresponds to a row in the table, and usually
there is only one FVector instance - the one we're searching for.
And there's an FVectorNode class, which is a node in the graph.
It has a ref (identifying a row in the source table), possibly an array
of floats (or not — in which case it will be read lazily from the
source table as needed). There are many FVectorNodes and they're
cached to avoid re-reading them from the disk.
instead of pointers to FVectorRef's (which are stored elsewhere)
let's return one big array of all refs. Freeing this array will
free the complete result set.
* sysvars should be REQUIRED_ARG
* fix a mix of US and UK spelling (use US)
* use consistent naming
* work if VEC_DISTANCE arguments are in the swapped order (const, col)
* work if VEC_DISTANCE argument is NULL/invalid or wrong length
* abort INSERT if the value is invalid or wrong length
* store the "number of neighbors" in a blob in endianness-independent way
* use field->store(longlong, bool) not field->store(double)
* a lot more error checking everywhere
* cleanup after errors
* simplify calling conventions, remove reinterpret_cast's
* todo/XXX comments
* whitespaces
* use float consistently
memory management is still totally PoC quality
This commit includes the work done in collaboration with Hugo Wen from
Amazon:
MDEV-33408 Alter HNSW graph storage and fix memory leak
This commit changes the way HNSW graph information is stored in the
second table. Instead of storing connections as separate records, it now
stores neighbors for each node, leading to significant performance
improvements and storage savings.
Comparing with the previous approach, the insert speed is 5 times faster,
search speed improves by 23%, and storage usage is reduced by 73%, based
on ann-benchmark tests with random-xs-20-euclidean and
random-s-100-euclidean datasets.
Additionally, in previous code, vector objects were not released after
use, resulting in excessive memory consumption (over 20GB for building
the index with 90,000 records), preventing tests with large datasets.
Now ensure that vectors are released appropriately during the insert and
search functions. Note there are still some vectors that need to be
cleaned up after search query completion. Needs to be addressed in a
future commit.
All new code of the whole pull request, including one or several files
that are either new files or modified ones, are contributed under the
BSD-new license. I am contributing on behalf of my employer Amazon Web
Services, Inc.
As well as the commit:
Introduce session variables to manage HNSW index parameters
Three variables:
hnsw_max_connection_per_layer
hnsw_ef_constructor
hnsw_ef_search
ann-benchmark tool is also updated to support these variables in commit
https://github.com/HugoWenTD/ann-benchmarks/commit/e09784e for branch
https://github.com/HugoWenTD/ann-benchmarks/tree/mariadb-configurable
All new code of the whole pull request, including one or several files
that are either new files or modified ones, are contributed under the
BSD-new license. I am contributing on behalf of my employer Amazon Web
Services, Inc.
Co-authored-by: Hugo Wen <wenhug@amazon.com>
MDEV-33407 Parser support for vector indexes
The syntax is
create table t1 (... vector index (v) ...);
limitation:
* v is a binary string and NOT NULL
* only one vector index per table
* temporary tables are not supported
MDEV-33404 Engine-independent indexes: subtable method
added support for so-called "high level indexes", they are not visible
to the storage engine, implemented on the sql level. For every such
an index in a table, say, t1, the server implicitly creates a second
table named, like, t1#i#05 (where "05" is the index number in t1).
This table has a fixed structure, no frm, not accessible directly,
doesn't go into the table cache, needs no MDLs.
MDEV-33406 basic optimizer support for k-NN searches
for a query like SELECT ... ORDER BY func() optimizer will use
item_func->part_of_sortkey() to decide what keys can be used
to resolve ORDER BY.
