Print oldest celldb snapshot to stats (#1078)

Co-authored-by: SpyCheese <mikle98@yandex.ru>

tddb/td/db/KeyValue.h

@@ -18,6 +18,7 @@
 */
 #pragma once
 #include "td/utils/Status.h"
+#include "td/utils/Time.h"
 #include "td/utils/logging.h"
 #include <functional>
 namespace td {

tddb/td/db/RocksDb.cpp

@@ -258,11 +258,17 @@ Status RocksDb::flush() {
 
 Status RocksDb::begin_snapshot() {
   snapshot_.reset(db_->GetSnapshot());
+  if (options_.snapshot_statistics) {
+    options_.snapshot_statistics->begin_snapshot(snapshot_.get());
+  }
   return td::Status::OK();
 }
 
 Status RocksDb::end_snapshot() {
   if (snapshot_) {
+    if (options_.snapshot_statistics) {
+      options_.snapshot_statistics->end_snapshot(snapshot_.get());
+    }
     db_->ReleaseSnapshot(snapshot_.release());
   }
   return td::Status::OK();
@@ -271,4 +277,41 @@ Status RocksDb::end_snapshot() {
 RocksDb::RocksDb(std::shared_ptr<rocksdb::OptimisticTransactionDB> db, RocksDbOptions options)
     : db_(std::move(db)), options_(options) {
 }
+
+void RocksDbSnapshotStatistics::begin_snapshot(const rocksdb::Snapshot *snapshot) {
+  auto lock = std::unique_lock<std::mutex>(mutex_);
+  auto id = reinterpret_cast<std::uintptr_t>(snapshot);
+  auto ts = td::Timestamp::now().at();
+  CHECK(id_to_ts_.emplace(id, ts).second);
+  CHECK(by_ts_.emplace(ts, id).second);
+}
+
+void RocksDbSnapshotStatistics::end_snapshot(const rocksdb::Snapshot *snapshot) {
+  auto id = reinterpret_cast<std::uintptr_t>(snapshot);
+  auto it = id_to_ts_.find(id);
+  CHECK(it != id_to_ts_.end());
+  auto ts = it->second;
+  CHECK(by_ts_.erase(std::make_pair(ts, id)) == 1u);
+  CHECK(id_to_ts_.erase(id) == 1u);
+}
+
+td::Timestamp RocksDbSnapshotStatistics::oldest_snapshot_timestamp() const {
+  auto lock = std::unique_lock<std::mutex>(mutex_);
+  if (by_ts_.empty()) {
+    return {};
+  }
+  return td::Timestamp::at(by_ts_.begin()->first);
+}
+
+std::string RocksDbSnapshotStatistics::to_string() const {
+  td::Timestamp oldest_snapshot = oldest_snapshot_timestamp();
+  double value;
+  if (oldest_snapshot) {
+    value = td::Timestamp::now().at() - oldest_snapshot.at();
+  } else {
+    value = -1;
+  }
+  return PSTRING() << "td.rocksdb.snapshot.oldest_snapshot_ago.seconds : " << value << "\n";
+}
+
 }  // namespace td

tddb/td/db/RocksDb.h

@@ -26,6 +26,12 @@
 #include "td/utils/Status.h"
 #include "td/utils/optional.h"
 
+#include "td/utils/Time.h"
+
+#include <map>
+#include <mutex>
+#include <set>
+
 namespace rocksdb {
 class Cache;
 class OptimisticTransactionDB;
@@ -36,10 +42,22 @@ class Statistics;
 }  // namespace rocksdb
 
 namespace td {
+struct RocksDbSnapshotStatistics {
+  void begin_snapshot(const rocksdb::Snapshot *snapshot);
+  void end_snapshot(const rocksdb::Snapshot *snapshot);
+  td::Timestamp oldest_snapshot_timestamp() const;
+  std::string to_string() const;
+
+ private:
+  mutable std::mutex mutex_;
+  std::map<std::uintptr_t, double> id_to_ts_;
+  std::set<std::pair<double, std::uintptr_t>> by_ts_;
+};
 
 struct RocksDbOptions {
   std::shared_ptr<rocksdb::Statistics> statistics = nullptr;
   std::shared_ptr<rocksdb::Cache> block_cache;  // Default - one 1G cache for all RocksDb
+  std::shared_ptr<RocksDbSnapshotStatistics> snapshot_statistics = nullptr;
   bool use_direct_reads = false;
 };
 

tddb/test/key_value.cpp

@@ -60,9 +60,19 @@ TEST(KeyValue, simple) {
   ensure_value(as_slice(x), as_slice(x));
 
   kv.reset();
-  kv = std::make_unique<td::RocksDb>(td::RocksDb::open(db_name.str()).move_as_ok());
+  td::RocksDbOptions options{.snapshot_statistics = std::make_shared<td::RocksDbSnapshotStatistics>()};
+  kv = std::make_unique<td::RocksDb>(td::RocksDb::open(db_name.str(), options).move_as_ok());
   ensure_value("A", "HELLO");
   ensure_value(as_slice(x), as_slice(x));
+
+  CHECK(!options.snapshot_statistics->oldest_snapshot_timestamp());
+  auto snapshot = kv->snapshot();
+  CHECK(options.snapshot_statistics->oldest_snapshot_timestamp());
+  auto snapshot2 = kv->snapshot();
+  snapshot.reset();
+  CHECK(options.snapshot_statistics->oldest_snapshot_timestamp());
+  snapshot2.reset();
+  CHECK(!options.snapshot_statistics->oldest_snapshot_timestamp());
 };
 
 TEST(KeyValue, async_simple) {

validator/db/celldb.cpp

@@ -84,11 +84,13 @@ void CellDbIn::start_up() {
   };
 
   CellDbBase::start_up();
+  td::RocksDbOptions db_options;
   if (!opts_->get_disable_rocksdb_stats()) {
     statistics_ = td::RocksDb::create_statistics();
     statistics_flush_at_ = td::Timestamp::in(60.0);
+    snapshot_statistics_ = std::make_shared<td::RocksDbSnapshotStatistics>();
+    db_options.snapshot_statistics = snapshot_statistics_;
   }
-  td::RocksDbOptions db_options;
   db_options.statistics = statistics_;
   if (opts_->get_celldb_cache_size()) {
     db_options.block_cache = td::RocksDb::create_cache(opts_->get_celldb_cache_size().value());
@@ -193,7 +195,11 @@ void CellDbIn::get_cell_db_reader(td::Promise<std::shared_ptr<vm::CellDbReader>>
 }
 
 void CellDbIn::flush_db_stats() {
-  auto stats = td::RocksDb::statistics_to_string(statistics_) + cell_db_statistics_.to_string();
+  if (opts_->get_disable_rocksdb_stats()) {
+    return;
+  }
+  auto stats = td::RocksDb::statistics_to_string(statistics_) + snapshot_statistics_->to_string() +
+               cell_db_statistics_.to_string();
   auto to_file_r =
       td::FileFd::open(path_ + "/db_stats.txt", td::FileFd::Truncate | td::FileFd::Create | td::FileFd::Write, 0644);
   if (to_file_r.is_error()) {

validator/db/celldb.hpp

@@ -27,6 +27,7 @@
 #include "auto/tl/ton_api.h"
 #include "validator.h"
 #include "db-utils.h"
+#include "td/db/RocksDb.h"
 
 namespace rocksdb {
 class Statistics;
@@ -139,6 +140,7 @@ class CellDbIn : public CellDbBase {
   };
 
   std::shared_ptr<rocksdb::Statistics> statistics_;
+  std::shared_ptr<td::RocksDbSnapshotStatistics> snapshot_statistics_;
   CellDbStatistics cell_db_statistics_;
   td::Timestamp statistics_flush_at_ = td::Timestamp::never();