Optimize LIKE with custom escape char (#7730)

Summary: Currently we optimize LIKE operation only if escape char is not specified, this PR adds the ability to apply the optimization even if user specifies escape char. We introduced a PatternStringIterator which handles escaping transparently, so existing optimizations(kPrefix, kSuffix, kSubstring etc) now work for patterns with escape char transparently, and future optimizations will have effect for escaped pattern transparently too. The benchmark result before this optimization: ``` ============================================================================ [...]hmarks/ExpressionBenchmarkBuilder.cpp relative time/iter iters/s ============================================================================ like_generic##like_generic 4.14s 241.44m ---------------------------------------------------------------------------- ---------------------------------------------------------------------------- like_prefix##like_prefix 1.20s 833.70m like_prefix##starts_with 2.92ms 342.44 like_substring##like_substring 4.22s 236.77m like_substring##strpos 6.98ms 143.27 like_suffix##like_suffix 3.09s 323.90m like_suffix##ends_with 3.02ms 331.11 ``` After: ``` ============================================================================ [...]hmarks/ExpressionBenchmarkBuilder.cpp relative time/iter iters/s ============================================================================ like_generic##like_generic 3.86s 258.97m ---------------------------------------------------------------------------- ---------------------------------------------------------------------------- like_prefix##like_prefix 4.18ms 239.24 like_prefix##starts_with 2.76ms 362.05 like_substring##like_substring 7.71ms 129.75 like_substring##strpos 6.67ms 149.90 like_suffix##like_suffix 4.20ms 237.85 like_suffix##ends_with 2.90ms 344.93 ``` In Summary: - Speedup of kSubstring is about 500x. - Speedup of kPrefix is about 250x. - Speedup of kSuffix is about 700x. Why the speedup is so huge? There are two reasons: - Re2 is really slow compare to the optimizations we made, even if the input string is short(10 byte), Re2 is 100x slower than our optimizations. - When the input strings get longer(10bytes -> 1000bytes), the performance of our optimizations does not change much, but Re2's performance will be 10x slower. And we can confirm the speedup is reasonable from the comparison between our optmizations and the simple scalar function strpos, starts_with, ends_with, the performance numbers are quite close(see the like##strpos/starts_with/ends_with in the benchmark result for more details). Pull Request resolved: #7730 Reviewed By: pedroerp Differential Revision: D52077250 Pulled By: mbasmanova fbshipit-source-id: 39703ddcc7f4f2044460d93866670f730e139120
facebookincubator · Dec 13, 2023 · 1779e82 · 1779e82
1 parent ddde53b
commit 1779e82
Show file tree

Hide file tree

Showing 6 changed files with 503 additions and 139 deletions.
diff --git a/velox/benchmarks/basic/CMakeLists.txt b/velox/benchmarks/basic/CMakeLists.txt
@@ -63,10 +63,15 @@ add_executable(velox_benchmark_basic_preproc Preproc.cpp)
 target_link_libraries(velox_benchmark_basic_preproc ${velox_benchmark_deps}
                       velox_functions_prestosql velox_vector_test_lib)
 
-add_executable(velox_like_functions_benchmark LikeFunctionsBenchmark.cpp)
-target_link_libraries(velox_like_functions_benchmark ${velox_benchmark_deps}
+add_executable(velox_like_tpch_benchmark LikeTpchBenchmark.cpp)
+target_link_libraries(velox_like_tpch_benchmark ${velox_benchmark_deps}
                       velox_functions_lib velox_tpch_gen velox_vector_test_lib)
 
+add_executable(velox_like_benchmark LikeBenchmark.cpp)
+target_link_libraries(
+  velox_like_benchmark ${velox_benchmark_deps} velox_functions_lib
+  velox_functions_prestosql velox_vector_test_lib)
+
 add_executable(velox_benchmark_basic_vector_fuzzer VectorFuzzer.cpp)
 target_link_libraries(velox_benchmark_basic_vector_fuzzer
                       ${velox_benchmark_deps} velox_vector_test_lib)

diff --git a/velox/benchmarks/basic/LikeBenchmark.cpp b/velox/benchmarks/basic/LikeBenchmark.cpp
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <folly/Benchmark.h>
+#include <folly/init/Init.h>
+
+#include "velox/benchmarks/ExpressionBenchmarkBuilder.h"
+#include "velox/functions/lib/Re2Functions.h"
+#include "velox/functions/prestosql/registration/RegistrationFunctions.h"
+
+using namespace facebook;
+using namespace facebook::velox;
+using namespace facebook::velox::functions;
+using namespace facebook::velox::functions::test;
+using namespace facebook::velox::memory;
+using namespace facebook::velox;
+
+int main(int argc, char** argv) {
+  folly::Init init(&argc, &argv);
+
+  exec::registerStatefulVectorFunction("like", likeSignatures(), makeLike);
+  // Register the scalar functions.
+  prestosql::registerAllScalarFunctions("");
+
+  // exec::register
+  ExpressionBenchmarkBuilder benchmarkBuilder;
+  const vector_size_t vectorSize = 1000;
+  auto vectorMaker = benchmarkBuilder.vectorMaker();
+
+  auto makeInput =
+      [&](vector_size_t vectorSize, bool padAtHead, bool padAtTail) {
+        return vectorMaker.flatVector<std::string>(vectorSize, [&](auto row) {
+          // Strings in even rows contain/start with/end with a_b_c depends on
+          // value of padAtHead && padAtTail.
+          if (row % 2 == 0) {
+            auto padding = std::string(row / 2 + 1, 'x');
+            if (padAtHead && padAtTail) {
+              return fmt::format("{}a_b_c{}", padding, padding);
+            } else if (padAtHead) {
+              return fmt::format("{}a_b_c", padding);
+            } else if (padAtTail) {
+              return fmt::format("a_b_c{}", padding);
+            } else {
+              return std::string("a_b_c");
+            }
+          } else {
+            return std::string(row, 'x');
+          }
+        });
+      };
+
+  auto substringInput = makeInput(vectorSize, true, true);
+  auto prefixInput = makeInput(vectorSize, false, true);
+  auto suffixInput = makeInput(vectorSize, true, false);
+
+  benchmarkBuilder
+      .addBenchmarkSet(
+          "like_substring", vectorMaker.rowVector({"col0"}, {substringInput}))
+      .addExpression("like_substring", R"(like(col0, '%a\_b\_c%', '\'))")
+      .addExpression("strpos", R"(strpos(col0, 'a_b_c') > 0)");
+
+  benchmarkBuilder
+      .addBenchmarkSet(
+          "like_prefix", vectorMaker.rowVector({"col0"}, {prefixInput}))
+      .addExpression("like_prefix", R"(like(col0, 'a\_b\_c%', '\'))")
+      .addExpression("starts_with", R"(starts_with(col0, 'a_b_c'))");
+
+  benchmarkBuilder
+      .addBenchmarkSet(
+          "like_suffix", vectorMaker.rowVector({"col0"}, {suffixInput}))
+      .addExpression("like_suffix", R"(like(col0, '%a\_b\_c', '\'))")
+      .addExpression("ends_with", R"(ends_with(col0, 'a_b_c'))");
+
+  benchmarkBuilder
+      .addBenchmarkSet(
+          "like_generic", vectorMaker.rowVector({"col0"}, {substringInput}))
+      .addExpression("like_generic", R"(like(col0, '%a%b%c'))");
+
+  benchmarkBuilder.registerBenchmarks();
+  benchmarkBuilder.testBenchmarks();
+  folly::runBenchmarks();
+  return 0;
+}
diff --git a/...nchmarks/basic/LikeFunctionsBenchmark.cpp → velox/benchmarks/basic/LikeTpchBenchmark.cpp b/...nchmarks/basic/LikeFunctionsBenchmark.cpp → velox/benchmarks/basic/LikeTpchBenchmark.cpp