- The patterns are not used in hashing, there should not be a performance
impact when we use stuff from the standard library instead.
- added forgotten .reserve() in FunctionsMultiStringPosition.h
This commit migrates ClickHouse to Vectorscan. The first 10 min of
[0] explain the reasons for it.
(*) Addresses (but does not resolve) #38046
(*) Config parameter names (e.g. "max_hyperscan_regexp_length") are
preserved for compatibility. Likewise, error codes (e.g.
"ErrorCodes::HYPERSCAN_CANNOT_SCAN_TEXT") and function/class names (e.g.
"HyperscanDeleter") are preserved as vectorscan aims to be a drop-in
replacement.
[0] https://www.youtube.com/watch?v=KlZWmmflW6M
First part, updated most UTF8, hashing, memory and codecs. Except
utf8lower and upper, maybe a little later.
That includes huge amount of research with movemask dealing. Exact
details and blog post TBD.
cmake/target.cmake defines macros for the supported platforms, this
commit changes predefined system macros to our own macros.
__linux__ --> OS_LINUX
__APPLE__ --> OS_DARWIN
__FreeBSD__ --> OS_FREEBSD
Trailing escape ('ab\') is disallowed in SQL, in standardese:
"If an escape character is specified, then [...] If there is not a
partitioning of the string PVC into substrings such that each substring
has length 1 (one) or 2, no substring of length 1 (one) is the escape
character ECV, and each substring of length 2 is the escape character
ECV followed by either the escape character ECV, an <underscore>
character, or the <percent> character, then an exception condition is
raised: data exception - invalid escape sequence."
I first thought this is checked already higher up in the stack, at least
for const needles, as single trailing backslashes ('ab\') are rejected,
but then I realized that ClickHouse quotes by default. I.e., double
trailing backslashes ('ab\\') are not rejected but when interpreted as
LIKE needle ('ab\') they should.
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.
Needles in a (non-const) needle column may repeat and this commit allows
to skip compilation for known needles. Out of the different design
alternatives (see below, if someone is interested), we now maintain
- one global pattern cache,
- with a fixed size of 42k elements currently,
- and use LRU as eviction strategy.
------------------------------------------------------------------------
(sorry for the wall of text, dumping it here not for reading but just
for reference)
Write-up about considered design alternatives:
1. Keep the current global cache of const needles. For non-const
needles, probe the cache but don't store values in it.
Pros: need to maintain just a single cache, no problem with cache
pollution assuming there are few distinct constant needles
Cons: only useful if a non-const needle occurred as already as a
const needle
--> overall too simplistic
2. Keep the current global cache for const needles. For non-const
needles, create a local (e.g. per-query) cache
Pros: unlike (1.), non-const needles can be skipped even if they
did not occur yet, no pollution of the const pattern cache when
there are very many non-const needles (e.g. large / highly
distinct needle columns).
Cons: caches may explode "horizontally", i.e. we'll end up with the
const cache + caches for Q1, Q2, ... QN, this makes it harder
to control the overall space consumption, also patterns
residing in different caches cannot be reused between queries,
another difficulty is that the concept of "query" does not
really exist at matching level - there are only column chunks
and we'd potentially end up with 1 cache / chunk
3. Queries with const and non-const needles insert into the same global
cache.
Pros: the advantages of (2.) + allows to reuse compiled patterns
accross parallel queries
Cons: needs an eviction strategy to control cache size and pollution
(and btw. (2.) also needs eviction strategies for the
individual caches)
4. Queries with const needle use global cache, queries with non-const
needle use a different global cache
--> Overall similar to (3) but ignores the (likely) edge case that
const and non-const needles overlap.
In sum, (3.) seems the simplest and most beneficial approach.
Eviction strategies:
0. Don't ever evict --> cache may grow infinitely and eventually make
the system unusable (may even pose a DoS risk)
1. Flush the cache after a certain threshold is exceeded --> very
simple but may lead to peridic performance drops
2. Use LRU --> more graceful performance degradation at threshold but
comes with a (constant) performance overhead to maintain the LRU
queue
In sum, given that the pattern compilation in RE2 should be quite costly
(pattern-to-DFA/NFA), LRU may be acceptable.
