Add options and code to disable all preloading of .sst files on DB::O…

…pen() (#33)
stardog-union · Aug 20, 2021 · d583bfc · d583bfc
1 parent bad4c41
commit d583bfc
Show file tree

Hide file tree

Showing 6 changed files with 52 additions and 11 deletions.
diff --git a/db/version_builder.cc b/db/version_builder.cc
@@ -63,7 +63,10 @@ class VersionBuilder::Rep {
   // kLevel0 -- NewestFirstBySeqNo
   // kLevelNon0 -- BySmallestKey
   struct FileComparator {
-    enum SortMethod { kLevel0 = 0, kLevelNon0 = 1, } sort_method;
+    enum SortMethod {
+      kLevel0 = 0,
+      kLevelNon0 = 1,
+    } sort_method;
     const InternalKeyComparator* internal_comparator;
 
     FileComparator() : internal_comparator(nullptr) {}
@@ -880,7 +883,7 @@ class VersionBuilder::Rep {
       auto added_end = added_files.end();
       while (added_iter != added_end || base_iter != base_end) {
         if (base_iter == base_end ||
-                (added_iter != added_end && cmp(*added_iter, *base_iter))) {
+            (added_iter != added_end && cmp(*added_iter, *base_iter))) {
           MaybeAddFile(vstorage, level, *added_iter++);
         } else {
           MaybeAddFile(vstorage, level, *base_iter++);
@@ -901,11 +904,15 @@ class VersionBuilder::Rep {
                            size_t max_file_size_for_l0_meta_pin) {
     assert(table_cache_ != nullptr);
 
+    if (kFilePreloadDisabled == ioptions_->file_preload) {
+      return Status::OK();
+    }
     size_t table_cache_capacity = table_cache_->get_cache()->GetCapacity();
     bool always_load = (table_cache_capacity == TableCache::kInfiniteCapacity);
     size_t max_load = port::kMaxSizet;
 #ifndef NDEBUG
-    bool debug_override = true;  // to enable CompactedDB related tests and some property tests
+    bool debug_override =
+        true;  // to enable CompactedDB related tests and some property tests
 #else
     bool debug_override = false;
 #endif
@@ -989,9 +996,8 @@ class VersionBuilder::Rep {
           } else {
             table_cache_->ReleaseHandle(file_meta->table_reader_handle);
             file_meta->table_reader_handle = nullptr;
-          } // else
+          }  // else
         }
-
       }
     });
 

diff --git a/db/version_edit_handler.cc b/db/version_edit_handler.cc
@@ -477,7 +477,8 @@ Status VersionEditHandler::MaybeCreateVersion(const VersionEdit& /*edit*/,
       // Install new version
       v->PrepareApply(
           *cfd->GetLatestMutableCFOptions(),
-          !(version_set_->db_options_->skip_stats_update_on_db_open));
+          !(version_set_->db_options_->skip_stats_update_on_db_open ||
+            kFilePreloadDisabled == cfd->ioptions()->file_preload));
       version_set_->AppendVersion(cfd, v);
     } else {
       delete v;
@@ -667,7 +668,8 @@ Status VersionEditHandlerPointInTime::MaybeCreateVersion(
     if (s.ok()) {
       version->PrepareApply(
           *cfd->GetLatestMutableCFOptions(),
-          !version_set_->db_options_->skip_stats_update_on_db_open);
+          !(version_set_->db_options_->skip_stats_update_on_db_open ||
+            kFilePreloadDisabled == cfd->ioptions()->file_preload));
       auto v_iter = versions_.find(cfd->GetID());
       if (v_iter != versions_.end()) {
         delete v_iter->second;

diff --git a/include/rocksdb/advanced_options.h b/include/rocksdb/advanced_options.h
@@ -56,6 +56,25 @@ enum CompactionPri : char {
   kMinOverlappingRatio = 0x3,
 };
 
+// RocksDB uses the first 25% of num_open_files for precaching during
+//  start-up and after compactions.  The files precached in this fashion
+//  provide faster access.  However, these files are also never released.
+//  Scenarios that have large bloom filters not cached or scenarios where
+//  user is manually lowering the num_open_files at runtime might want
+//  to disable this behavior.
+enum FilePreload : char {
+  // RocksDB uses the first 25% of num_open_files for precaching during
+  //  start-up and after compactions.  The files precached in this fashion
+  //  provide faster access.  However, these files are also never released.
+  kFilePreloadWithPinning = 0x0,
+  // RocksDB uses the first 25% of num_open_files for precaching during
+  //  start-up and after compactions.  No pinning within cache, so access
+  //  has one additional layer of indirection.  But cache space can free.
+  kFilePreloadWithoutPinning = 0x1,
+  // RocksDB does not open existing table files during start-up.
+  kFilePreloadDisabled = 0x2,
+};
+
 struct CompactionOptionsFIFO {
   // once the total sum of table files reaches this, we will delete the oldest
   // table file
@@ -691,6 +710,13 @@ struct AdvancedColumnFamilyOptions {
   // Default: true
   bool force_consistency_checks = true;
 
+  // RocksDB can preload and optionally pin table files within the table
+  //  cache at start-up and after compactions.  The files precached in this
+  //  fashion provide faster access.  However, these files are also never
+  //  released from the table cache.
+  // Default: kFilePreloadWithPinning
+  FilePreload file_preload = kFilePreloadWithPinning;
+
   // Measure IO stats in compactions and flushes, if true.
   //
   // Default: false

diff --git a/include/rocksdb/write_batch.h b/include/rocksdb/write_batch.h
@@ -179,7 +179,7 @@ class WriteBatch : public WriteBatchBase {
   Status PopSavePoint() override;
 
   void setContentFlag(uint32_t theContentFlag) {
-    content_flags_.store(theContentFlag,std::memory_order_seq_cst);
+    content_flags_.store(theContentFlag, std::memory_order_seq_cst);
   }
 
   // Support for iterating over the contents of a batch.

diff --git a/options/cf_options.cc b/options/cf_options.cc
@@ -491,6 +491,10 @@ static std::unordered_map<std::string, OptionTypeInfo>
         {"compaction_measure_io_stats",
          {0, OptionType::kBoolean, OptionVerificationType::kDeprecated,
           OptionTypeFlags::kNone}},
+        {"file_preload",
+         {offset_of(&ColumnFamilyOptions::file_preload),
+          OptionType::kFilePreload, OptionVerificationType::kNormal,
+          OptionTypeFlags::kNone}},
         {"inplace_update_support",
          {offset_of(&ColumnFamilyOptions::inplace_update_support),
           OptionType::kBoolean, OptionVerificationType::kNormal,
@@ -841,6 +845,7 @@ ImmutableCFOptions::ImmutableCFOptions(const ImmutableDBOptions& db_options,
       num_levels(cf_options.num_levels),
       optimize_filters_for_hits(cf_options.optimize_filters_for_hits),
       force_consistency_checks(cf_options.force_consistency_checks),
+      file_preload(cf_options.file_preload),
       allow_ingest_behind(db_options.allow_ingest_behind),
       preserve_deletes(db_options.preserve_deletes),
       listeners(db_options.listeners),
@@ -869,9 +874,9 @@ uint64_t MultiplyCheckOverflow(uint64_t op1, double op2) {
 // when level_compaction_dynamic_level_bytes is true and leveled compaction
 // is used, the base level is not always L1, so precomupted max_file_size can
 // no longer be used. Recompute file_size_for_level from base level.
-uint64_t MaxFileSizeForLevel(const MutableCFOptions& cf_options,
-    int level, CompactionStyle compaction_style, int base_level,
-    bool level_compaction_dynamic_level_bytes) {
+uint64_t MaxFileSizeForLevel(const MutableCFOptions& cf_options, int level,
+                             CompactionStyle compaction_style, int base_level,
+                             bool level_compaction_dynamic_level_bytes) {
   if (!level_compaction_dynamic_level_bytes || level < base_level ||
       compaction_style != kCompactionStyleLevel) {
     assert(level >= 0);

diff --git a/options/cf_options.h b/options/cf_options.h
@@ -105,6 +105,8 @@ struct ImmutableCFOptions {
 
   bool force_consistency_checks;
 
+  FilePreload file_preload;
+
   bool allow_ingest_behind;
 
   bool preserve_deletes;