Adding more processors for parallelize_output_from_storages is not a
costless operation (I've experienced some issues in production because
of this), and it is not easy to fix in a normal way, so let's disable it
for now.
Before this patch:
- INSERT INTO input SELECT * FROM numbers(10e6) SETTINGS parallelize_output_from_storages=1, min_insert_block_size_rows=1000
0 rows in set. Elapsed: 3.648 sec. Processed 20.00 million rows, 120.00 MB (5.48 million rows/s., 32.90 MB/s.)
- INSERT INTO input SELECT * FROM numbers(10e6) SETTINGS parallelize_output_from_storages=0, min_insert_block_size_rows=1000
0 rows in set. Elapsed: 1.851 sec. Processed 20.00 million rows, 120.00 MB (10.80 million rows/s., 64.82 MB/s.)
Signed-off-by: Azat Khuzhin <a.khuzhin@semrush.com>
As it turns out, HashMap/PackedHashMap works great even with max load
factor of 0.99. By "great" I mean it least it works faster then
google sparsehash, and not to mention it's friendliness to the memory
allocator (it has zero fragmentation since it works with a continuious
memory region, in comparison to the sparsehash that doing lots of
realloc, which jemalloc does not like, due to it's slabs).
Here is a table of different setups:
settings | load (sec) | read (sec) | read (million rows/s) | bytes_allocated | RSS
- | - | - | - | - | -
HASHED upstream | - | - | - | - | 35GiB
SPARSE_HASHED upstream | - | - | - | - | 26GiB
- | - | - | - | - | -
sparse_hash_map glibc hashbench | - | - | - | - | 17.5GiB
sparse_hash_map packed allocator | 101.878 | 231.48 | 4.32 | - | 17.7GiB
PackedHashMap 0.5 | 15.514 | 42.35 | 23.61 | 20GiB | 22GiB
hashed 0.95 | 34.903 | 115.615 | 8.65 | 16GiB | 18.7GiB
**PackedHashMap 0.95** | **93.6** | **19.883** | **10.68** | **10GiB** | **12.8GiB**
PackedHashMap 0.99 | 26.113 | 83.6 | 11.96 | 10GiB | 12.3GiB
As it shows, PackedHashMap with 0.95 max_load_factor, eats 2.6x less
memory then SPARSE_HASHED in upstream, and it also 2x faster for read!
v2: fix grower
Signed-off-by: Azat Khuzhin <a.khuzhin@semrush.com>
In case you want dictionary optimized for memory, SPARSE_HASHED is not
always gives you what you need.
Consider the following example <UInt64, UInt16> as <Key, Value>, but
this pair will also have a 6 byte padding (on amd64), so this is almost
40% of space wastage.
And because of this padding, even google::sparse_hash_map, does not make
picture better, in fact, sparse_hash_map is not very friendly to memory
allocators (especially jemalloc).
Here are some numbers for dictionary with 1e9 elements and UInt64 as
key, and UInt16 as value:
settings | load (sec) | read (sec) | read (million rows/s) | bytes_allocated | RSS
HASHED upstream | - | - | - | - | 35GiB
SPARSE_HASHED upstream | - | - | - | - | 26GiB
- | - | - | - | - | -
sparse_hash_map glibc hashbench | - | - | - | - | 17.5GiB
sparse_hash_map packed allocator | 101.878 | 231.48 | 4.32 | - | 17.7GiB
PackedHashMap | 15.514 | 42.35 | 23.61 | 20GiB | 22GiB
As you can see PackedHashMap looks way more better then HASHED, and
even better then SPARSE_HASHED, but slightly worse then sparse_hash_map
with packed allocator (it is done with a custom patch to google
sparse_hash_map).
v2: rebase on top of bucket_count fix
Signed-off-by: Azat Khuzhin <a.khuzhin@semrush.com>
Reasons:
1. The original Gorilla paper proposed a compression schema for pairs of
time stamps and double-precision FP values. ClickHouse's Gorilla
codec only implements compression of the latter and it does not
impose any data type restrictions.
- Data types != Float* or (U)Int* (e.g. Decimal, Point etc.) are
definitely not supposed to be used with Gorilla.
- (U)Int* types are debatable. The paper only considers
integers-stored-as-FP-values, a practical use case for which
Gorilla works well. Standalone integers are not considered which
makes them at least suspicious.
2. Achieve consistency with FPC, another specialized floating-point
timeseries codec, which rejects non-float data.
3. On practical datasets, ZSTD is often "good enough" (**) so it should
be okay to disincentive non-ZSTD codecs a little bit. If needed,
Delta and DoubleDelta codecs are viable alternative for slowly
changing (time-series-like) integer sequences.
Since on-prem and hosted users may still have Gorilla-compressed
non-float data, this combination is only deprecated for now. No warning
or error will be emitted. Users are encouraged to migrate
Gorilla-compressed non-float data to an alternative codec. It is planned
to treat Gorilla-compressed non-float columns as "suspicious" six months
after this commit (i.e. in v23.6). Even then, it will still be possible
to set "allow_suspicious_codecs = true" and read and write
Gorilla-compressed non-float data.
(*) Sec. 4.1.2, "Gorilla restricts the value element in its tuple to a
double floating point type.", https://doi.org/10.14778/2824032.2824078
(**) https://clickhouse.com/blog/optimize-clickhouse-codecs-compression-schema
Right now dictionaries (here I will talk about only
HASHED/SPARSE_HASHED/COMPLEX_KEY_HASHED/COMPLEX_KEY_SPARSE_HASHED)
can load data only in one thread, since it uses one hash table that
cannot be filled from multiple threads.
And in case you have very big dictionary (i.e. 10e9 elements), it can
take a awhile to load them, especially for SPARSE_HASHED variants (and
if you have such amount of elements there, you are likely use
SPARSE_HASHED, since it requires less memory), in my env it takes ~4
hours, which is enormous amount of time.
So this patch add support of shards for dictionaries, number of shards
determine how much hash tables will use this dictionary, also, and which
is more important, how much threads it can use to load the data.
And with 16 threads this works 2x faster, not perfect though, see the
follow up patches in this series.
v0: PARTITION BY
v1: SHARDS 1
v2: SHARDS(1)
v3: tried optimized mod - logical and, but it does not gain even 10%
v4: tried squashing more (max_block_size * shards), but it does not gain even 10% either
v5: move SHARDS into layout parameters (unknown simply ignored)
v6: tune params for perf tests (to avoid too long queries)
Signed-off-by: Azat Khuzhin <a.khuzhin@semrush.com>
* impl
* fix style
* make executeQueryWithParallelReplicas similar to executeQuery
* impl for parallel replicas
* cleaner code for remote sorting properties
* update test
* fix
* handle when nodes of old versions participate
* small fixes
* temporary enable for testing
* fix after merge
* Revert "temporary enable for testing"
This reverts commit cce7f8884c.
* review fixes
* add bc test
* Update src/Core/Settings.h
* impl
* stash
* clean up
* do not apply when HT is small
* make branch static
* also in merge
* do not hardcode look ahead value
* fix
* apply to methods with cheap key calculation
* more tests
* silence tidy
* fix build
* support HashMethodKeysFixed
* apply during merge only for cheap
* stash
* fixes
* rename method
* add feature flag
* cache prefetch threshold value
* fix
* fix
* Update HashMap.h
* fix typo
* 256KB as default l2 size
Co-authored-by: Alexey Milovidov <milovidov@clickhouse.com>
+ range search done in steps of some number of rows.
Controled by new
setting `distinct_in_order_range_search_step`. By default 0, i.e.
whole chunk
+ before start binary search, linear probing is done on each step (32
rows currently)
This commit is based on local benchmarks of ClickHouse's re2 caching.
Question 1: -----------------------------------------------------------
Is pattern caching useful for queries with const LIKE/REGEX
patterns? E.g. SELECT LIKE(col_haystack, '%HelloWorld') FROM T;
The short answer is: no. Runtime is (unsurprisingly) dominated by
pattern evaluation + other stuff going on in queries, but definitely not
pattern compilation. For space reasons, I omit details of the local
experiments.
(Side note: the current caching scheme is unbounded in size which poses
a DoS risk (think of multi-tenancy). This risk is more pronounced when
unbounded caching is used with non-const patterns ..., see next
question)
Question 2: -----------------------------------------------------------
Is pattern caching useful for queries with non-const LIKE/REGEX
patterns? E.g. SELECT LIKE(col_haystack, col_needle) FROM T;
I benchmarked five caching strategies:
1. no caching as a baseline (= recompile for each row)
2. unbounded cache (= threadsafe global hash-map)
3. LRU cache (= threadsafe global hash-map + LRU queue)
4. lightweight local cache 1 (= not threadsafe local hashmap with
collision list which grows to a certain size (here: 10 elements) and
afterwards never changes)
5. lightweight local cache 2 (not threadsafe local hashmap without
collision list in which a collision replaces the stored element, idea
by Alexey)
... using a haystack of 2 mio strings and
A). 2 mio distinct simple patterns
B). 10 simple patterns
C) 2 mio distinct complex patterns
D) 10 complex patterns
Fo A) and C), caching does not help but these queries still allow to
judge the static overhead of caching on query runtimes.
B) and D) are extreme but common cases in practice. They include
queries like "SELECT ... WHERE LIKE (col_haystack, flag ? '%pattern1%' :
'%pattern2%'). Caching should help significantly.
Because LIKE patterns are internally translated to re2 expressions, I
show only measurements for MATCH queries.
Results in sec, averaged over on multiple measurements;
1.A): 2.12
B): 1.68
C): 9.75
D): 9.45
2.A): 2.17
B): 1.73
C): 9.78
D): 9.47
3.A): 9.8
B): 0.63
C): 31.8
D): 0.98
4.A): 2.14
B): 0.29
C): 9.82
D): 0.41
5.A) 2.12 / 2.15 / 2.26
B) 1.51 / 0.43 / 0.30
C) 9.97 / 9.88 / 10.13
D) 5.70 / 0.42 / 0.43
(10/100/1000 buckets, resp. 10/1/0.1% collision rate)
Evaluation:
1. This is the baseline. It was surprised that complex patterns (C, D)
slow down the queries so badly compared to simple patterns (A, B).
The runtime includes evaluation costs, but as caching only helps with
compilation, and looking at 4.D and 5.D, compilation makes up over 90%
of the runtime!
2. No speedup compared to 1, probably due to locking overhead. The cache
is unbounded, and in experiments with data sets > 2 mio rows, 2. is
the only scheme to throw OOM exceptions which is not acceptable.
3. Unique patterns (A and C) lead to thrashing of the LRU cache and very
bad runtimes due to LRU queue maintenance and locking. Works pretty
well however with few distinct patterns (B and D).
4. This scheme is tailored to queries B and D where it performs pretty
good. More importantly, the caching is lightweight enough to not
deteriorate performance on datasets A and C.
5. After some tuning of the hash map size, 100 buckets seem optimal to
be in the same ballpark with 10 distinct patterns as 4. Performance
also does not deteriorate on A and C compared to the baseline.
Unlike 4., this scheme behaves LRU-like and can adjust to changing
pattern distributions.
As a conclusion, this commit implementes two things:
1. Based on Q1, pattern search with const needle no longer uses
caching. This applies to LIKE and MATCH + a few (exotic) other SQL
functions. The code for the unbounded caching was removed.
2. Based on Q2, pattern search with non-const needles now use method 5.
