check last sample time when gc

pika/HISTORY.md

@@ -1,4 +1,8 @@
 # pika for codis
+## 3.0.4-3.2 (2019-10-23)
+### bug fix
+* 修复服务重启后，每个blob的可回收率的分数被清为0，只有被访问到的sst文件中的blob数据会重新计算，导致很多blob文件分数虽然高，但采样时又不满足回收条件的文件一直被选中，后面的无效数据很多但分数为0的真正需要回收的blob文件无法被删除，从而导致gc速度就变慢的bug
+
 ## 3.0.4-3.1 (2019-10-10)
 ### New Features
 * 周期检测操作系统free内存剩余大小，如果小于用户设定值，则清理。只有master时执行该操作

pika/include/pika_version.h

@@ -10,6 +10,6 @@
 #define PIKA_MINOR 0
 #define PIKA_PATCH 4
 #define PIKA_XMLY_MAJOR  3
-#define PIKA_XMLY_MINOR  1
+#define PIKA_XMLY_MINOR  2
 
 #endif  // INCLUDE_PIKA_VERSION_H_

pika/third/rocksdb/utilities/titandb/blob_format.h

@@ -1,5 +1,7 @@
 #pragma once
 
+#include <atomic>
+
 #include "rocksdb/options.h"
 #include "rocksdb/slice.h"
 #include "rocksdb/status.h"
@@ -140,6 +142,9 @@ class BlobFileMeta {
   void AddDiscardableSize(uint64_t _discardable_size);
   double GetDiscardableRatio() const;
 
+  int64_t last_sample_time() const { return last_sample_time_; }
+  void set_last_sample_time(int64_t timestamp) { last_sample_time_ = timestamp; }
+
  private:
   // Persistent field
   uint64_t file_number_{0};
@@ -150,6 +155,9 @@ class BlobFileMeta {
 
   uint64_t discardable_size_{0};
   //  bool marked_for_gc_{false};
+
+  // last gc sample time
+  std::atomic<int64_t> last_sample_time_{0};
 };
 
 // Blob file footer format:

pika/third/rocksdb/utilities/titandb/blob_gc_job.cc

@@ -98,6 +98,12 @@ Status BlobGCJob::Run() {
       tmp.append(" ");
     }
     tmp.append(std::to_string(f->file_number()));
+    tmp.append("(");
+    std::ostringstream score;
+    score.precision(2);
+    score << f->GetDiscardableRatio();
+    tmp.append(score.str());
+    tmp.append(")");
   }
 
   std::string tmp2;
@@ -122,7 +128,7 @@ Status BlobGCJob::Run() {
 
 Status BlobGCJob::SampleCandidateFiles() {
   std::vector<BlobFileMeta*> result;
-  for (const auto& file : blob_gc_->inputs()) {
+  for (auto& file : blob_gc_->inputs()) {
     if (DoSample(file)) {
       result.push_back(file);
     }
@@ -135,7 +141,7 @@ Status BlobGCJob::SampleCandidateFiles() {
   return Status::OK();
 }
 
-bool BlobGCJob::DoSample(const BlobFileMeta* file) {
+bool BlobGCJob::DoSample(BlobFileMeta* file) {
   if (file->GetDiscardableRatio() >=
       blob_gc_->titan_cf_options().blob_file_discardable_ratio) {
     return true;
@@ -149,6 +155,11 @@ bool BlobGCJob::DoSample(const BlobFileMeta* file) {
     return true;
   }
 
+  // update file sample time
+  int64_t unix_time;
+  Env::Default()->GetCurrentTime(&unix_time);
+  file->set_last_sample_time(unix_time);
+
   Status s;
   uint64_t sample_size_window = static_cast<uint64_t>(
       file->file_size() * blob_gc_->titan_cf_options().sample_file_size_ratio);

pika/third/rocksdb/utilities/titandb/blob_gc_job.h

@@ -58,7 +58,7 @@ class BlobGCJob {
   std::atomic_bool* shuting_down_{nullptr};
 
   Status SampleCandidateFiles();
-  bool DoSample(const BlobFileMeta* file);
+  bool DoSample(BlobFileMeta* file);
   Status DoRunGC();
   Status BuildIterator(std::unique_ptr<BlobFileMergeIterator>* result);
   bool DiscardEntry(const Slice& key, const BlobIndex& blob_index);

pika/third/rocksdb/utilities/titandb/blob_gc_picker.cc

@@ -3,6 +3,8 @@
 namespace rocksdb {
 namespace titandb {
 
+static const int64_t ONE_DAY_SECONDS = 86400;
+
 BasicBlobGCPicker::BasicBlobGCPicker(TitanDBOptions db_options,
                                      TitanCFOptions cf_options)
     : db_options_(db_options), cf_options_(cf_options) {}
@@ -17,6 +19,10 @@ std::unique_ptr<BlobGC> BasicBlobGCPicker::PickBlobGC(
   uint64_t batch_size = 0;
 //  ROCKS_LOG_INFO(db_options_.info_log, "blob file num:%lu gc score:%lu",
 //                 blob_storage->NumBlobFiles(), blob_storage->gc_score().size());
+
+  int64_t unix_time;
+  Env::Default()->GetCurrentTime(&unix_time);
+
   for (auto& gc_score : blob_storage->gc_score()) {
     auto blob_file = blob_storage->FindFile(gc_score.file_number).lock();
     assert(blob_file);
@@ -36,12 +42,30 @@ std::unique_ptr<BlobGC> BasicBlobGCPicker::PickBlobGC(
                      blob_file->file_number());
       continue;
     }
+
+    // if the file has sampled last time, but not gc, we will skip the file
+    if (blob_file->GetDiscardableRatio() < cf_options_.blob_file_discardable_ratio) {
+      if (0 != blob_file->last_sample_time() 
+          && unix_time - blob_file->last_sample_time() < ONE_DAY_SECONDS) {
+          ROCKS_LOG_DEBUG(db_options_.info_log, "Titan GC skip the file[%lu]",
+                          gc_score.file_number);
+          continue;
+      }
+    }
+
     blob_files.push_back(blob_file.get());
 
     batch_size += blob_file->file_size();
     if (batch_size >= cf_options_.max_gc_batch_size) break;
   }
 
+  // reset file last_sample_time if we have checked out to the end of all blob files
+  if (blob_files.empty()) {
+    ROCKS_LOG_INFO(db_options_.info_log, "Titan GC check to the end of all blob files, blob file num:%lu gc score:%lu",
+                   blob_storage->NumBlobFiles(), blob_storage->gc_score().size());
+    ResetAllBlobFileSampleTime(blob_storage);
+  }
+
   if (blob_files.empty() || batch_size < cf_options_.min_gc_batch_size)
     return nullptr;
 
@@ -56,5 +80,16 @@ bool BasicBlobGCPicker::CheckBlobFile(BlobFileMeta* blob_file) const {
   return true;
 }
 
+void BasicBlobGCPicker::ResetAllBlobFileSampleTime(BlobStorage* blob_storage) {
+  for (auto& gc_score : blob_storage->gc_score()) {
+    auto blob_file = blob_storage->FindFile(gc_score.file_number).lock();
+    assert(blob_file);
+    if (!CheckBlobFile(blob_file.get())) {
+      continue;
+    }
+    blob_file->set_last_sample_time(0);
+  }
+}
+
 }  // namespace titandb
 }  // namespace rocksdb

pika/third/rocksdb/utilities/titandb/blob_gc_picker.h

@@ -40,6 +40,8 @@ class BasicBlobGCPicker final : public BlobGCPicker {
   // Check if blob_file needs to gc, return true means we need pick this
   // file for gc
   bool CheckBlobFile(BlobFileMeta* blob_file) const;
+
+  void ResetAllBlobFileSampleTime(BlobStorage* blob_storage);
 };
 
 }  // namespace titandb