Add punch hole GC

Signed-off-by: v01dstar <[email protected]>

include/titan/db.h

@@ -201,6 +201,9 @@ class TitanDB : public StackableDB {
     //  "rocksdb.titandb.discardable_ratio_le100_file_num" - returns count of
     //      file whose discardable ratio is less or equal to 100%.
     static const std::string kNumDiscardableRatioLE100File;
+    //  "rockdb.titandb.kNumHolePunchableBlobSize" - returns the size of hole
+    //      punchable blobs (no longer referenced in SSTs) in the database.
+    static const std::string kHolePunchableBlobSize;
   };
 
   bool GetProperty(ColumnFamilyHandle* column_family, const Slice& property,

include/titan/options.h

@@ -172,6 +172,7 @@ struct TitanCFOptions : public ColumnFamilyOptions {
   // data's 0s and 0s created by punch holes).
   uint64_t block_size{4096};
   bool enable_punch_hole_gc{false};
+  uint64_t punch_hole_threshold{4 * 1024 * 1024};
 
   TitanCFOptions() = default;
   explicit TitanCFOptions(const ColumnFamilyOptions& options)
@@ -229,12 +230,14 @@ struct MutableTitanCFOptions {
       : blob_run_mode(opts.blob_run_mode),
         min_blob_size(opts.min_blob_size),
         blob_file_compression(opts.blob_file_compression),
-        blob_file_discardable_ratio(opts.blob_file_discardable_ratio) {}
+        blob_file_discardable_ratio(opts.blob_file_discardable_ratio),
+        punch_hole_threshold(opts.punch_hole_threshold) {}
 
   TitanBlobRunMode blob_run_mode;
   uint64_t min_blob_size;
   CompressionType blob_file_compression;
   double blob_file_discardable_ratio;
+  uint64_t punch_hole_threshold;
 };
 
 struct TitanOptions : public TitanDBOptions, public TitanCFOptions {

src/blob_file_builder.cc

@@ -164,7 +164,12 @@ void BlobFileBuilder::FlushSampleRecords(OutContexts* out_ctx) {
 void BlobFileBuilder::WriteEncoderData(BlobHandle* handle) {
   handle->offset = file_->GetFileSize();
   handle->size = encoder_.GetEncodedSize();
-  live_data_size_ += handle->size;
+  if (block_size_ > 0) {
+    live_data_size_ +=
+        (handle->size + block_size_ - 1) / block_size_ * block_size_;
+  } else {
+    live_data_size_ += handle->size;
+  }
 
   status_ = file_->Append(encoder_.GetHeader());
   if (ok()) {

src/blob_file_builder.h

@@ -108,6 +108,8 @@ class BlobFileBuilder {
   const std::string& GetSmallestKey() { return smallest_key_; }
   const std::string& GetLargestKey() { return largest_key_; }
 
+  uint64_t GetBlockSize() { return block_size_; }
+
   uint64_t live_data_size() const { return live_data_size_; }
 
  private:

src/blob_file_manager.h

@@ -72,6 +72,15 @@ class BlobFileManager {
     (void)handles;
     return Status::OK();
   }
+
+  // Updates the metadata of the file. This is used to update the
+  // metadata of the file after the file is punched with holes.
+  virtual Status BatchUpdateFiles(
+      uint32_t cf_id, const std::vector<std::shared_ptr<BlobFileMeta>>& files) {
+    (void)cf_id;
+    (void)files;
+    return Status::OK();
+  }
 };
 
 }  // namespace titandb

src/blob_file_set.h

@@ -98,6 +98,22 @@ class BlobFileSet {
 
   bool IsOpened() { return opened_.load(std::memory_order_acquire); }
 
+  uint64_t GetBlockSize(uint32_t cf_id) {
+    MutexLock l(mutex_);
+    auto storage = GetBlobStorage(cf_id).lock();
+    if (storage != nullptr && storage->cf_options().enable_punch_hole_gc) {
+      return storage->cf_options().block_size;
+    }
+    return 0;
+  }
+
+  std::unordered_map<uint64_t, uint64_t> GetFileBlockSizes(uint32_t cf_id) {
+    MutexLock l(mutex_);
+    auto storage = GetBlobStorage(cf_id).lock();
+    return storage ? storage->GetFileBlockSizes()
+                   : std::unordered_map<uint64_t, uint64_t>();
+  }
+
  private:
   struct ManifestWriter;
 

src/blob_file_size_collector.cc

@@ -7,8 +7,13 @@ namespace titandb {
 
 TablePropertiesCollector*
 BlobFileSizeCollectorFactory::CreateTablePropertiesCollector(
-    rocksdb::TablePropertiesCollectorFactory::Context /* context */) {
-  return new BlobFileSizeCollector();
+    rocksdb::TablePropertiesCollectorFactory::Context context) {
+  if (blob_file_set_ != nullptr) {
+    return new BlobFileSizeCollector(
+        blob_file_set_->GetBlockSize(context.column_family_id),
+        blob_file_set_->GetFileBlockSizes(context.column_family_id));
+  }
+  return new BlobFileSizeCollector(0, {});
 }
 
 const std::string BlobFileSizeCollector::kPropertiesName =
@@ -57,11 +62,32 @@ Status BlobFileSizeCollector::AddUserKey(const Slice& /* key */,
     return s;
   }
 
+  auto size = index.blob_handle.size;
+  if (default_block_size_ > 0 && !file_block_sizes_.empty()) {
+    // If the blob file cannot be found in the block size map, it must be a
+    // newly created file that has not been added blob_file_set, in this case,
+    // we know the block size of the file is default_block_size_.
+    // If the blob file can be found in the block size map, it implies we are
+    // moving the reference only, while keeping the blob at the original file,
+    // in this case, we should use the block size of the original file.
+    uint64_t block_size = default_block_size_;
+    if (!file_block_sizes_.empty()) {
+      auto iter = file_block_sizes_.find(index.file_number);
+      if (iter != file_block_sizes_.end()) {
+        block_size = iter->second;
+      }
+    }
+    if (block_size > 0) {
+      // Align blob size with block size.
+      size = (size + block_size - 1) / block_size * block_size;
+    }
+  }
+
   auto iter = blob_files_size_.find(index.file_number);
   if (iter == blob_files_size_.end()) {
-    blob_files_size_[index.file_number] = index.blob_handle.size;
+    blob_files_size_[index.file_number] = size;
   } else {
-    iter->second += index.blob_handle.size;
+    iter->second += size;
   }
 
   return Status::OK();

src/blob_file_size_collector.h

@@ -13,16 +13,31 @@ namespace titandb {
 class BlobFileSizeCollectorFactory final
     : public TablePropertiesCollectorFactory {
  public:
+  // If punch_hole_gc is enabled, then blob_file_set must be provided.
+  // If blob_file_set is not provided, then punch_hole_gc will be considered
+  // disabled, blob size will not align with block size.
+  BlobFileSizeCollectorFactory(BlobFileSet* blob_file_set = nullptr)
+      : blob_file_set_(blob_file_set) {}
+  BlobFileSizeCollectorFactory(const BlobFileSizeCollectorFactory&) = delete;
+  void operator=(const BlobFileSizeCollectorFactory&) = delete;
   TablePropertiesCollector* CreateTablePropertiesCollector(
       TablePropertiesCollectorFactory::Context context) override;
 
   const char* Name() const override { return "BlobFileSizeCollector"; }
+
+ private:
+  BlobFileSet* blob_file_set_;
 };
 
 class BlobFileSizeCollector final : public TablePropertiesCollector {
  public:
   const static std::string kPropertiesName;
 
+  BlobFileSizeCollector(uint64_t default_block_size,
+                        std::unordered_map<uint64_t, uint64_t> file_block_sizes)
+      : default_block_size_(default_block_size),
+        file_block_sizes_(file_block_sizes) {}
+
   static bool Encode(const std::map<uint64_t, uint64_t>& blob_files_size,
                      std::string* result);
   static bool Decode(Slice* slice,
@@ -38,6 +53,8 @@ class BlobFileSizeCollector final : public TablePropertiesCollector {
 
  private:
   std::map<uint64_t, uint64_t> blob_files_size_;
+  uint64_t default_block_size_;
+  std::unordered_map<uint64_t, uint64_t> file_block_sizes_;
 };
 
 }  // namespace titandb

src/blob_format.cc

@@ -142,6 +142,7 @@ void BlobFileMeta::EncodeTo(std::string* dst) const {
   PutVarint64(dst, block_size_);
   PutLengthPrefixedSlice(dst, smallest_key_);
   PutLengthPrefixedSlice(dst, largest_key_);
+  PutVarint64(dst, effective_file_size_);
 }
 
 Status BlobFileMeta::DecodeFromV1(Slice* src) {
@@ -189,6 +190,12 @@ Status BlobFileMeta::DecodeFrom(Slice* src) {
   } else {
     return Status::Corruption("BlobLargestKey decode failed");
   }
+  uint64_t effective_file_size;
+  if (!GetVarint64(src, &effective_file_size)) {
+    return Status::Corruption(
+        "BlobFileMeta hole_punchable_size_ decode failed");
+  }
+  effective_file_size_ = effective_file_size;
   return Status::OK();
 }
 

src/blob_format.h

@@ -248,10 +248,14 @@ class BlobFileMeta {
   uint64_t file_size() const { return file_size_; }
   uint64_t live_data_size() const { return live_data_size_; }
   uint32_t file_level() const { return file_level_; }
+  uint64_t block_size() const { return block_size_; }
   const std::string& smallest_key() const { return smallest_key_; }
   const std::string& largest_key() const { return largest_key_; }
+  int64_t effective_file_size() const { return effective_file_size_; }
 
   void set_live_data_size(int64_t size) { live_data_size_ = size; }
+  // This should be called with db mutex held.
+  void set_effective_file_size(int64_t size) { effective_file_size_ = size; }
   uint64_t file_entries() const { return file_entries_; }
   FileState file_state() const { return state_; }
   bool is_obsolete() const { return state_ == FileState::kObsolete; }
@@ -275,6 +279,10 @@ class BlobFileMeta {
                 (file_size_ - kBlobMaxHeaderSize - kBlobFooterSize));
   }
   TitanInternalStats::StatsType GetDiscardableRatioLevel() const;
+  // This should be called with db mutex held.
+  uint64_t GetHolePunchableSize() const {
+    return effective_file_size_ - live_data_size_;
+  }
   void Dump(bool with_keys) const;
 
  private:
@@ -291,6 +299,18 @@ class BlobFileMeta {
   std::string smallest_key_;
   std::string largest_key_;
 
+  // The effective size of current file. This is different from `file_size_`, as
+  // `file_size_` is the original size of the file, and does not consider space
+  // reclaimed by punch hole GC.
+  // We can't use file system's `st_blocks` to get the logical size, because
+  // the file system's block size may be different from Titan's block size.
+  // This is used to calculate the size of the punchable hole. i.e.
+  // effective_file_size_ - live_data_size_.
+  // This might be bigger than the actual size of the file, when Titan crashes
+  // before updating the `effective_file_size_` during punch hole GC. This is
+  // fine, as it will be corrected when the file is chose for GC next time.
+  int64_t effective_file_size_{0};
+
   // Not persistent field
 
   // Size of data with reference from SST files.
@@ -303,7 +323,11 @@ class BlobFileMeta {
   // So when state_ == kPendingLSM, it uses this to record the delta as a
   // positive number if any later compaction is trigger before previous
   // `OnCompactionCompleted()` is called.
+  // The size is aligned with block size, when punch hole GC is enabled.
   std::atomic<int64_t> live_data_size_{0};
+  // This is different from `file_size_`, as `file_size_` is the original size
+  // of the file, and does not consider space reclaimed by punch hole GC.
+  std::atomic<int64_t> disk_usage_{0};
   std::atomic<FileState> state_{FileState::kNone};
 };
 

src/blob_gc.cc

@@ -4,10 +4,12 @@ namespace rocksdb {
 namespace titandb {
 
 BlobGC::BlobGC(std::vector<std::shared_ptr<BlobFileMeta>>&& blob_files,
-               TitanCFOptions&& _titan_cf_options, bool need_trigger_next)
+               TitanCFOptions&& _titan_cf_options, bool need_trigger_next,
+               bool punch_hole_gc)
     : inputs_(blob_files),
       titan_cf_options_(std::move(_titan_cf_options)),
-      trigger_next_(need_trigger_next) {
+      trigger_next_(need_trigger_next),
+      punch_hole_gc_(punch_hole_gc) {
   MarkFilesBeingGC();
 }
 

src/blob_gc.h

@@ -14,7 +14,8 @@ namespace titandb {
 class BlobGC {
  public:
   BlobGC(std::vector<std::shared_ptr<BlobFileMeta>>&& blob_files,
-         TitanCFOptions&& _titan_cf_options, bool need_trigger_next);
+         TitanCFOptions&& _titan_cf_options, bool need_trigger_next,
+         bool punch_hole_gc = false);
 
   // No copying allowed
   BlobGC(const BlobGC&) = delete;
@@ -40,13 +41,16 @@ class BlobGC {
 
   bool trigger_next() { return trigger_next_; }
 
+  bool punch_hole_gc() { return punch_hole_gc_; }
+
  private:
   std::vector<std::shared_ptr<BlobFileMeta>> inputs_;
   std::vector<BlobFileMeta*> outputs_;
   TitanCFOptions titan_cf_options_;
   ColumnFamilyHandle* cfh_{nullptr};
   // Whether need to trigger gc after this gc or not
   const bool trigger_next_;
+  const bool punch_hole_gc_;
 };
 
 struct GCScore {

src/blob_gc_picker.cc

@@ -18,7 +18,19 @@ BasicBlobGCPicker::BasicBlobGCPicker(TitanDBOptions db_options,
 
 BasicBlobGCPicker::~BasicBlobGCPicker() {}
 
-std::unique_ptr<BlobGC> BasicBlobGCPicker::PickBlobGC(
+std::unique_ptr<BlobGC> BasicBlobGCPicker::PickBlobGC(BlobStorage* blob_storage,
+                                                      bool allow_punch_hole) {
+  auto regular_gc = PickRegularBlobGC(blob_storage);
+  if (regular_gc) {
+    return regular_gc;
+  }
+  if (allow_punch_hole) {
+    return PickPunchHoleGC(blob_storage);
+  }
+  return nullptr;
+}
+
+std::unique_ptr<BlobGC> BasicBlobGCPicker::PickRegularBlobGC(
     BlobStorage* blob_storage) {
   Status s;
   std::vector<std::shared_ptr<BlobFileMeta>> blob_files;
@@ -103,6 +115,47 @@ std::unique_ptr<BlobGC> BasicBlobGCPicker::PickBlobGC(
       std::move(blob_files), std::move(cf_options_), maybe_continue_next_time));
 }
 
+std::unique_ptr<BlobGC> BasicBlobGCPicker::PickPunchHoleGC(
+    BlobStorage* blob_storage) {
+  Status s;
+  std::vector<std::shared_ptr<BlobFileMeta>> blob_files;
+
+  uint64_t batch_size = 0;
+  uint64_t estimate_output_size = 0;
+  bool stop_picking = false;
+  bool maybe_continue_next_time = false;
+  uint64_t next_gc_size = 0;
+
+  for (auto& gc_score : blob_storage->punch_hole_score()) {
+    auto blob_file = blob_storage->FindFile(gc_score.file_number).lock();
+    if (!CheckBlobFile(blob_file.get())) {
+      // Skip this file id this file is being GCed
+      // or this file had been GCed
+      TITAN_LOG_INFO(db_options_.info_log,
+                     "Blob file %" PRIu64 " no need punch hole gc",
+                     blob_file->file_number());
+      continue;
+    }
+    if (!stop_picking) {
+      blob_files.emplace_back(blob_file);
+      batch_size += blob_file->file_size();
+      if (batch_size >= cf_options_.max_gc_batch_size) {
+        // Stop pick file for this gc, but still check file for whether need
+        // trigger gc after this
+        stop_picking = true;
+      }
+    } else {
+      // TODO: add a batch threshold for punch hole gc.
+      maybe_continue_next_time = true;
+      break;
+    }
+  }
+  if (blob_files.empty()) return nullptr;
+  return std::unique_ptr<BlobGC>(new BlobGC(
+      std::move(blob_files), std::move(cf_options_), maybe_continue_next_time,
+      /*punch_hole_gc=*/true));
+}
+
 bool BasicBlobGCPicker::CheckBlobFile(BlobFileMeta* blob_file) const {
   assert(blob_file == nullptr ||
          blob_file->file_state() != BlobFileMeta::FileState::kNone);