pass key/value samples through zstd compression dictionary generator

Summary:
Instead of using samples directly, we now support passing the samples through zstd's dictionary generator when `CompressionOptions::zstd_max_train_bytes` is set to nonzero. If set to zero, we will use the samples directly as the dictionary -- same as before.

Note this is the first step of #2987, extracted into a separate PR per reviewer request.
Closes https://github.com/facebook/rocksdb/pull/3057

Differential Revision: D6116891

Pulled By: ajkr

fbshipit-source-id: 70ab13cc4c734fa02e554180eed0618b75255497

db/column_family.cc

@@ -123,6 +123,22 @@ Status CheckCompressionSupported(const ColumnFamilyOptions& cf_options) {
           " is not linked with the binary.");
     }
   }
+  if (cf_options.compression_opts.zstd_max_train_bytes > 0) {
+    if (!CompressionTypeSupported(CompressionType::kZSTD)) {
+      // Dictionary trainer is available since v0.6.1, but ZSTD was marked
+      // stable only since v0.8.0. For now we enable the feature in stable
+      // versions only.
+      return Status::InvalidArgument(
+          "zstd dictionary trainer cannot be used because " +
+          CompressionTypeToString(CompressionType::kZSTD) +
+          " is not linked with the binary.");
+    }
+    if (cf_options.compression_opts.max_dict_bytes == 0) {
+      return Status::InvalidArgument(
+          "The dictionary size limit (`CompressionOptions::max_dict_bytes`) "
+          "should be nonzero if we're using zstd's dictionary generator.");
+    }
+  }
   return Status::OK();
 }
 

db/compaction_job.cc

@@ -702,15 +702,18 @@ void CompactionJob::ProcessKeyValueCompaction(SubcompactionState* sub_compact) {
       sub_compact->compaction->mutable_cf_options();
 
   // To build compression dictionary, we sample the first output file, assuming
-  // it'll reach the maximum length, and then use the dictionary for compressing
-  // subsequent output files. The dictionary may be less than max_dict_bytes if
-  // the first output file's length is less than the maximum.
+  // it'll reach the maximum length. We optionally pass these samples through
+  // zstd's dictionary trainer, or just use them directly. Then, the dictionary
+  // is used for compressing subsequent output files in the same subcompaction.
+  const bool kUseZstdTrainer =
+      cfd->ioptions()->compression_opts.zstd_max_train_bytes > 0;
+  const size_t kSampleBytes =
+      kUseZstdTrainer ? cfd->ioptions()->compression_opts.zstd_max_train_bytes
+                      : cfd->ioptions()->compression_opts.max_dict_bytes;
   const int kSampleLenShift = 6;  // 2^6 = 64-byte samples
   std::set<size_t> sample_begin_offsets;
-  if (bottommost_level_ &&
-      cfd->ioptions()->compression_opts.max_dict_bytes > 0) {
-    const size_t kMaxSamples =
-        cfd->ioptions()->compression_opts.max_dict_bytes >> kSampleLenShift;
+  if (bottommost_level_ && kSampleBytes > 0) {
+    const size_t kMaxSamples = kSampleBytes >> kSampleLenShift;
     const size_t kOutFileLen = mutable_cf_options->MaxFileSizeForLevel(
         compact_->compaction->output_level());
     if (kOutFileLen != port::kMaxSizet) {
@@ -780,11 +783,11 @@ void CompactionJob::ProcessKeyValueCompaction(SubcompactionState* sub_compact) {
   }
   const auto& c_iter_stats = c_iter->iter_stats();
   auto sample_begin_offset_iter = sample_begin_offsets.cbegin();
-  // data_begin_offset and compression_dict are only valid while generating
+  // data_begin_offset and dict_sample_data are only valid while generating
   // dictionary from the first output file.
   size_t data_begin_offset = 0;
-  std::string compression_dict;
-  compression_dict.reserve(cfd->ioptions()->compression_opts.max_dict_bytes);
+  std::string dict_sample_data;
+  dict_sample_data.reserve(kSampleBytes);
 
   while (status.ok() && !cfd->IsDropped() && c_iter->Valid()) {
     // Invariant: c_iter.status() is guaranteed to be OK if c_iter->Valid()
@@ -856,7 +859,7 @@ void CompactionJob::ProcessKeyValueCompaction(SubcompactionState* sub_compact) {
             data_elmt_copy_len =
                 data_end_offset - (data_begin_offset + data_elmt_copy_offset);
           }
-          compression_dict.append(&data_elmt.data()[data_elmt_copy_offset],
+          dict_sample_data.append(&data_elmt.data()[data_elmt_copy_offset],
                                   data_elmt_copy_len);
           if (sample_end_offset > data_end_offset) {
             // Didn't finish sample. Try to finish it with the next data_elmt.
@@ -911,9 +914,15 @@ void CompactionJob::ProcessKeyValueCompaction(SubcompactionState* sub_compact) {
       RecordDroppedKeys(range_del_out_stats,
                         &sub_compact->compaction_job_stats);
       if (sub_compact->outputs.size() == 1) {
-        // Use dictionary from first output file for compression of subsequent
-        // files.
-        sub_compact->compression_dict = std::move(compression_dict);
+        // Use samples from first output file to create dictionary for
+        // compression of subsequent files.
+        if (kUseZstdTrainer) {
+          sub_compact->compression_dict = ZSTD_TrainDictionary(
+              dict_sample_data, kSampleLenShift,
+              cfd->ioptions()->compression_opts.max_dict_bytes);
+        } else {
+          sub_compact->compression_dict = std::move(dict_sample_data);
+        }
       }
     }
   }

db/db_test2.cc

@@ -1027,6 +1027,7 @@ TEST_F(DBTest2, PresetCompressionDict) {
   const size_t kL0FileBytes = 128 << 10;
   const size_t kApproxPerBlockOverheadBytes = 50;
   const int kNumL0Files = 5;
+  const int kZstdTrainFactor = 16;
 
   Options options;
   options.env = CurrentOptions().env; // Make sure to use any custom env that the test is configured with.
@@ -1059,17 +1060,34 @@ TEST_F(DBTest2, PresetCompressionDict) {
   for (auto compression_type : compression_types) {
     options.compression = compression_type;
     size_t prev_out_bytes;
-    for (int i = 0; i < 2; ++i) {
+    for (int i = 0; i < 3; ++i) {
       // First iteration: compress without preset dictionary
       // Second iteration: compress with preset dictionary
-      // To make sure the compression dictionary was actually used, we verify
-      // the compressed size is smaller in the second iteration. Also in the
-      // second iteration, verify the data we get out is the same data we put
-      // in.
-      if (i) {
-        options.compression_opts.max_dict_bytes = kBlockSizeBytes;
-      } else {
+      // Third iteration (zstd only): compress with zstd-trained dictionary
+      //
+      // To make sure the compression dictionary has the intended effect, we
+      // verify the compressed size is smaller in successive iterations. Also in
+      // the non-first iterations, verify the data we get out is the same data
+      // we put in.
+      switch (i) {
+        case 0:
           options.compression_opts.max_dict_bytes = 0;
+          options.compression_opts.zstd_max_train_bytes = 0;
+          break;
+        case 1:
+          options.compression_opts.max_dict_bytes = kBlockSizeBytes;
+          options.compression_opts.zstd_max_train_bytes = 0;
+          break;
+        case 2:
+          if (compression_type != kZSTD) {
+            continue;
+          }
+          options.compression_opts.max_dict_bytes = kBlockSizeBytes;
+          options.compression_opts.zstd_max_train_bytes =
+              kZstdTrainFactor * kBlockSizeBytes;
+          break;
+        default:
+          assert(false);
       }
 
       options.statistics = rocksdb::CreateDBStatistics();

include/rocksdb/advanced_options.h

@@ -90,23 +90,47 @@ struct CompressionOptions {
   int window_bits;
   int level;
   int strategy;
-  // Maximum size of dictionary used to prime the compression library. Currently
-  // this dictionary will be constructed by sampling the first output file in a
-  // subcompaction when the target level is bottommost. This dictionary will be
-  // loaded into the compression library before compressing/uncompressing each
-  // data block of subsequent files in the subcompaction. Effectively, this
-  // improves compression ratios when there are repetitions across data blocks.
-  // A value of 0 indicates the feature is disabled.
+
+  // Maximum size of dictionaries used to prime the compression library.
+  // Enabling dictionary can improve compression ratios when there are
+  // repetitions across data blocks.
+  //
+  // The dictionary is created by sampling the SST file data. If
+  // `zstd_max_train_bytes` is nonzero, the samples are passed through zstd's
+  // dictionary generator. Otherwise, the random samples are used directly as
+  // the dictionary.
+  //
+  // When compression dictionary is disabled, we compress and write each block
+  // before buffering data for the next one. When compression dictionary is
+  // enabled, we buffer all SST file data in-memory so we can sample it, as data
+  // can only be compressed and written after the dictionary has been finalized.
+  // So users of this feature may see increased memory usage.
+  //
   // Default: 0.
   uint32_t max_dict_bytes;
 
+  // Maximum size of training data passed to zstd's dictionary trainer. Using
+  // zstd's dictionary trainer can achieve even better compression ratio
+  // improvements than using `max_dict_bytes` alone.
+  //
+  // The training data will be used to generate a dictionary of max_dict_bytes.
+  //
+  // Default: 0.
+  uint32_t zstd_max_train_bytes;
+
   CompressionOptions()
-      : window_bits(-14), level(-1), strategy(0), max_dict_bytes(0) {}
-  CompressionOptions(int wbits, int _lev, int _strategy, int _max_dict_bytes)
+      : window_bits(-14),
+        level(-1),
+        strategy(0),
+        max_dict_bytes(0),
+        zstd_max_train_bytes(0) {}
+  CompressionOptions(int wbits, int _lev, int _strategy, int _max_dict_bytes,
+                     int _zstd_max_train_bytes)
       : window_bits(wbits),
         level(_lev),
         strategy(_strategy),
-        max_dict_bytes(_max_dict_bytes) {}
+        max_dict_bytes(_max_dict_bytes),
+        zstd_max_train_bytes(_zstd_max_train_bytes) {}
 };
 
 enum UpdateStatus {    // Return status For inplace update callback

util/compression.h

@@ -35,7 +35,10 @@
 
 #if defined(ZSTD)
 #include <zstd.h>
-#endif
+#if ZSTD_VERSION_NUMBER >= 800  // v0.8.0+
+#include <zdict.h>
+#endif  // ZSTD_VERSION_NUMBER >= 800
+#endif  // ZSTD
 
 #if defined(XPRESS)
 #include "port/xpress.h"
@@ -796,4 +799,42 @@ inline char* ZSTD_Uncompress(const char* input_data, size_t input_length,
   return nullptr;
 }
 
+inline std::string ZSTD_TrainDictionary(const std::string& samples,
+                                        const std::vector<size_t>& sample_lens,
+                                        size_t max_dict_bytes) {
+  // Dictionary trainer is available since v0.6.1, but ZSTD was marked stable
+  // only since v0.8.0. For now we enable the feature in stable versions only.
+#if ZSTD_VERSION_NUMBER >= 800  // v0.8.0+
+  std::string dict_data(max_dict_bytes, '\0');
+  size_t dict_len =
+      ZDICT_trainFromBuffer(&dict_data[0], max_dict_bytes, &samples[0],
+                            &sample_lens[0], sample_lens.size());
+  if (ZDICT_isError(dict_len)) {
+    return "";
+  }
+  assert(dict_len <= max_dict_bytes);
+  dict_data.resize(dict_len);
+  return dict_data;
+#else   // up to v0.7.x
+  assert(false);
+  return "";
+#endif  // ZSTD_VERSION_NUMBER >= 800
+}
+
+inline std::string ZSTD_TrainDictionary(const std::string& samples,
+                                        size_t sample_len_shift,
+                                        size_t max_dict_bytes) {
+  // Dictionary trainer is available since v0.6.1, but ZSTD was marked stable
+  // only since v0.8.0. For now we enable the feature in stable versions only.
+#if ZSTD_VERSION_NUMBER >= 800  // v0.8.0+
+  // skips potential partial sample at the end of "samples"
+  size_t num_samples = samples.size() >> sample_len_shift;
+  std::vector<size_t> sample_lens(num_samples, 1 << sample_len_shift);
+  return ZSTD_TrainDictionary(samples, sample_lens, max_dict_bytes);
+#else   // up to v0.7.x
+  assert(false);
+  return "";
+#endif  // ZSTD_VERSION_NUMBER >= 800
+}
+
 }  // namespace rocksdb