Add timeout_hint_us to WriteOptions and introduce Status::TimeOut.

Summary:
This diff adds timeout_hint_us to WriteOptions.  If it's non-zero, then
1) writes associated with this options MAY be aborted when it has been
  waiting for longer than the specified time.  If an abortion happens,
  associated writes will return Status::TimeOut.
2) the stall time of the associated write caused by flush or compaction
  will be limited by timeout_hint_us.

The default value of timeout_hint_us is 0 (i.e., OFF.)

The statistics of timeout writes will be recorded in WRITE_TIMEDOUT.

Test Plan:
export ROCKSDB_TESTS=WriteTimeoutAndDelayTest
make db_test
./db_test

Reviewers: igor, ljin, haobo, sdong

Reviewed By: sdong

Subscribers: dhruba, leveldb

Differential Revision: https://reviews.facebook.net/D18837

db/db_impl.cc

@@ -80,7 +80,9 @@ struct DBImpl::Writer {
   WriteBatch* batch;
   bool sync;
   bool disableWAL;
+  bool in_batch_group;
   bool done;
+  uint64_t timeout_hint_us;
   port::CondVar cv;
 
   explicit Writer(port::Mutex* mu) : cv(mu) { }
@@ -3729,13 +3731,41 @@ Status DBImpl::Write(const WriteOptions& options, WriteBatch* my_batch) {
   w.batch = my_batch;
   w.sync = options.sync;
   w.disableWAL = options.disableWAL;
+  w.in_batch_group = false;
+  w.done = false;
+  w.timeout_hint_us = options.timeout_hint_us;
+
+  uint64_t expiration_time = 0;
+  if (w.timeout_hint_us == 0) {
+    w.timeout_hint_us = kNoTimeOut;
+  } else {
+    expiration_time = env_->NowMicros() + w.timeout_hint_us;
+  }
   w.done = false;
 
-  StopWatch sw(env_, options_.statistics.get(), DB_WRITE, false);
   mutex_.Lock();
+  // the following code block pushes the current writer "w" into the writer
+  // queue "writers_" and wait until one of the following conditions met:
+  // 1. the job of "w" has been done by some other writers.
+  // 2. "w" becomes the first writer in "writers_"
+  // 3. "w" timed-out.
   writers_.push_back(&w);
+
+  bool timed_out = false;
   while (!w.done && &w != writers_.front()) {
+    if (expiration_time == 0) {
       w.cv.Wait();
+    } else if (w.cv.TimedWait(expiration_time)) {
+      if (w.in_batch_group) {
+        // then it means the front writer is currently doing the
+        // write on behalf of this "timed-out" writer.  Then it
+        // should wait until the write completes.
+        expiration_time = 0;
+      } else {
+        timed_out = true;
+        break;
+      }
+    }
   }
 
   if (!options.disableWAL) {
@@ -3746,10 +3776,33 @@ Status DBImpl::Write(const WriteOptions& options, WriteBatch* my_batch) {
     mutex_.Unlock();
     RecordTick(options_.statistics.get(), WRITE_DONE_BY_OTHER, 1);
     return w.status;
+  } else if (timed_out) {
+    bool found = false;
+    for (auto iter = writers_.begin(); iter != writers_.end(); iter++) {
+      if (*iter == &w) {
+        writers_.erase(iter);
+        found = true;
+        break;
+      }
+    }
+    assert(found);
+    // writers_.front() might still be in cond_wait without a time-out.
+    // As a result, we need to signal it to wake it up.  Otherwise no
+    // one else will wake him up, and RocksDB will hang.
+    if (!writers_.empty()) {
+      writers_.front()->cv.Signal();
+    }
+    mutex_.Unlock();
+    RecordTick(options_.statistics.get(), WRITE_TIMEDOUT, 1);
+    return Status::TimedOut();
   } else {
     RecordTick(options_.statistics.get(), WRITE_DONE_BY_SELF, 1);
   }
 
+  // Once reaches this point, the current writer "w" will try to do its write
+  // job.  It may also pick up some of the remaining writers in the "writers_"
+  // when it finds suitable, and finish them in the same write batch.
+  // This is how a write job could be done by the other writer.
   assert(!single_column_family_mode_ ||
          versions_->GetColumnFamilySet()->NumberOfColumnFamilies() == 1);
 
@@ -3774,8 +3827,10 @@ Status DBImpl::Write(const WriteOptions& options, WriteBatch* my_batch) {
 
   if (LIKELY(single_column_family_mode_)) {
     // fast path
-    status = MakeRoomForWrite(default_cf_handle_->cfd(), my_batch == nullptr,
-                              &superversions_to_free, &logs_to_free);
+    status = MakeRoomForWrite(
+        default_cf_handle_->cfd(), my_batch == nullptr,
+        &superversions_to_free, &logs_to_free,
+        expiration_time);
   } else {
     // refcounting cfd in iteration
     bool dead_cfd = false;
@@ -3786,8 +3841,9 @@ Status DBImpl::Write(const WriteOptions& options, WriteBatch* my_batch) {
           (flush_column_family_if_log_file != 0 &&
            cfd->GetLogNumber() <= flush_column_family_if_log_file);
       // May temporarily unlock and wait.
-      status = MakeRoomForWrite(cfd, force_flush, &superversions_to_free,
-                                &logs_to_free);
+      status = MakeRoomForWrite(
+          cfd, force_flush, &superversions_to_free, &logs_to_free,
+          expiration_time);
       if (cfd->Unref()) {
         dead_cfd = true;
       }
@@ -3883,11 +3939,14 @@ Status DBImpl::Write(const WriteOptions& options, WriteBatch* my_batch) {
       }
     }
   }
-  if (options_.paranoid_checks && !status.ok() && bg_error_.ok()) {
+  if (options_.paranoid_checks && !status.ok() &&
+      !status.IsTimedOut() && bg_error_.ok()) {
     bg_error_ = status; // stop compaction & fail any further writes
   }
 
-  while (true) {
+  // Pop out the current writer and all writers being pushed before the
+  // current writer from the writer queue.
+  while (!writers_.empty()) {
     Writer* ready = writers_.front();
     writers_.pop_front();
     if (ready != &w) {
@@ -3904,6 +3963,10 @@ Status DBImpl::Write(const WriteOptions& options, WriteBatch* my_batch) {
   }
   mutex_.Unlock();
 
+  if (status.IsTimedOut()) {
+    RecordTick(options_.statistics.get(), WRITE_TIMEDOUT, 1);
+  }
+
   for (auto& sv : superversions_to_free) {
     delete sv;
   }
@@ -3915,6 +3978,9 @@ Status DBImpl::Write(const WriteOptions& options, WriteBatch* my_batch) {
   return status;
 }
 
+// This function will be called only when the first writer succeeds.
+// All writers in the to-be-built batch group will be processed.
+//
 // REQUIRES: Writer list must be non-empty
 // REQUIRES: First writer must have a non-nullptr batch
 void DBImpl::BuildBatchGroup(Writer** last_writer,
@@ -3950,6 +4016,12 @@ void DBImpl::BuildBatchGroup(Writer** last_writer,
       break;
     }
 
+    if (w->timeout_hint_us < first->timeout_hint_us) {
+      // Do not include those writes with shorter timeout.  Otherwise, we might
+      // execute a write that should instead be aborted because of timeout.
+      break;
+    }
+
     if (w->batch != nullptr) {
       size += WriteBatchInternal::ByteSize(w->batch);
       if (size > max_size) {
@@ -3959,6 +4031,7 @@ void DBImpl::BuildBatchGroup(Writer** last_writer,
 
       write_batch_group->push_back(w->batch);
     }
+    w->in_batch_group = true;
     *last_writer = w;
   }
 }
@@ -4000,7 +4073,8 @@ uint64_t DBImpl::SlowdownAmount(int n, double bottom, double top) {
 Status DBImpl::MakeRoomForWrite(
     ColumnFamilyData* cfd, bool force,
     autovector<SuperVersion*>* superversions_to_free,
-    autovector<log::Writer*>* logs_to_free) {
+    autovector<log::Writer*>* logs_to_free,
+    uint64_t expiration_time) {
   mutex_.AssertHeld();
   assert(!writers_.empty());
   bool allow_delay = !force;
@@ -4013,12 +4087,16 @@ Status DBImpl::MakeRoomForWrite(
   // might generate a tight feedback loop, constantly scheduling more background
   // work, even if additional background work is not needed
   bool schedule_background_work = true;
+  bool has_timeout = (expiration_time > 0);
 
   while (true) {
     if (!bg_error_.ok()) {
       // Yield previous error
       s = bg_error_;
       break;
+    } else if (has_timeout && env_->NowMicros() > expiration_time) {
+      s = Status::TimedOut();
+      break;
     } else if (allow_delay && cfd->NeedSlowdownForNumLevel0Files()) {
       // We are getting close to hitting a hard limit on the number of
       // L0 files.  Rather than delaying a single write by several
@@ -4063,7 +4141,11 @@ Status DBImpl::MakeRoomForWrite(
       {
         StopWatch sw(env_, options_.statistics.get(),
                      STALL_MEMTABLE_COMPACTION_COUNT);
+        if (!has_timeout) {
           bg_cv_.Wait();
+        } else {
+          bg_cv_.TimedWait(expiration_time);
+        }
         stall = sw.ElapsedMicros();
       }
       RecordTick(options_.statistics.get(),
@@ -4078,10 +4160,15 @@ Status DBImpl::MakeRoomForWrite(
       {
         StopWatch sw(env_, options_.statistics.get(),
                      STALL_L0_NUM_FILES_COUNT);
+        if (!has_timeout) {
           bg_cv_.Wait();
+        } else {
+          bg_cv_.TimedWait(expiration_time);
+        }
         stall = sw.ElapsedMicros();
       }
-      RecordTick(options_.statistics.get(), STALL_L0_NUM_FILES_MICROS, stall);
+      RecordTick(options_.statistics.get(),
+                 STALL_L0_NUM_FILES_MICROS, stall);
       cfd->internal_stats()->RecordWriteStall(InternalStats::LEVEL0_NUM_FILES,
                                               stall);
     } else if (allow_hard_rate_limit_delay && cfd->ExceedsHardRateLimit()) {
@@ -4112,18 +4199,18 @@ Status DBImpl::MakeRoomForWrite(
       score = cfd->current()->MaxCompactionScore();
       // Delay a write when the compaction score for any level is too large.
       // TODO: add statistics
+      uint64_t slowdown = SlowdownAmount(score, cfd->options()->soft_rate_limit,
+                                         cfd->options()->hard_rate_limit);
       mutex_.Unlock();
       {
         StopWatch sw(env_, options_.statistics.get(),
                      SOFT_RATE_LIMIT_DELAY_COUNT);
-        env_->SleepForMicroseconds(
-            SlowdownAmount(score, cfd->options()->soft_rate_limit,
-                           cfd->options()->hard_rate_limit));
-        rate_limit_delay_millis += sw.ElapsedMicros();
+        env_->SleepForMicroseconds(slowdown);
+        slowdown = sw.ElapsedMicros();
+        rate_limit_delay_millis += slowdown;
       }
       allow_soft_rate_limit_delay = false;
       mutex_.Lock();
-
     } else {
       unique_ptr<WritableFile> lfile;
       log::Writer* new_log = nullptr;

db/db_impl.h

@@ -10,6 +10,7 @@
 
 #include <atomic>
 #include <deque>
+#include <limits>
 #include <set>
 #include <utility>
 #include <vector>
@@ -28,6 +29,7 @@
 #include "rocksdb/transaction_log.h"
 #include "util/autovector.h"
 #include "util/stats_logger.h"
+#include "util/stop_watch.h"
 #include "util/thread_local.h"
 #include "db/internal_stats.h"
 
@@ -345,7 +347,8 @@ class DBImpl : public DB {
   Status MakeRoomForWrite(ColumnFamilyData* cfd,
                           bool force /* flush even if there is room? */,
                           autovector<SuperVersion*>* superversions_to_free,
-                          autovector<log::Writer*>* logs_to_free);
+                          autovector<log::Writer*>* logs_to_free,
+                          uint64_t expiration_time);
 
   void BuildBatchGroup(Writer** last_writer,
                        autovector<WriteBatch*>* write_batch_group);
@@ -578,6 +581,7 @@ class DBImpl : public DB {
   bool flush_on_destroy_; // Used when disableWAL is true.
 
   static const int KEEP_LOG_FILE_NUM = 1000;
+  static const uint64_t kNoTimeOut = std::numeric_limits<uint64_t>::max();
   std::string db_absolute_path_;
 
   // count of the number of contiguous delaying writes

db/db_test.cc

@@ -7001,6 +7001,128 @@ TEST(DBTest, FIFOCompactionTest) {
     }
   }
 }
+
+TEST(DBTest, SimpleWriteTimeoutTest) {
+  Options options;
+  options.env = env_;
+  options.create_if_missing = true;
+  options.write_buffer_size = 100000;
+  options.max_background_flushes = 0;
+  options.max_write_buffer_number = 2;
+  options.min_write_buffer_number_to_merge = 3;
+  options.max_total_wal_size = std::numeric_limits<uint64_t>::max();
+  WriteOptions write_opt = WriteOptions();
+  write_opt.timeout_hint_us = 500;
+  DestroyAndReopen(&options);
+  // fill the two write buffer
+  ASSERT_OK(Put(Key(1), Key(1) + std::string(100000, 'v'), write_opt));
+  ASSERT_OK(Put(Key(2), Key(2) + std::string(100000, 'v'), write_opt));
+  // As the only two write buffers are full in this moment, the third
+  // Put is expected to be timed-out.
+  ASSERT_TRUE(
+      Put(Key(3), Key(3) + std::string(100000, 'v'), write_opt).IsTimedOut());
+}
+
+// Multi-threaded Timeout Test
+namespace {
+
+static const int kValueSize = 1000;
+static const int kWriteBufferSize = 100000;
+
+struct TimeoutWriterState {
+  int id;
+  DB* db;
+  std::atomic<bool> done;
+  std::map<int, std::string> success_kvs;
+};
+
+static void RandomTimeoutWriter(void* arg) {
+  TimeoutWriterState* state = reinterpret_cast<TimeoutWriterState*>(arg);
+  static const uint64_t kTimerBias = 50;
+  int thread_id = state->id;
+  DB* db = state->db;
+
+  Random rnd(1000 + thread_id);
+  WriteOptions write_opt = WriteOptions();
+  write_opt.timeout_hint_us = 500;
+  int timeout_count = 0;
+  int num_keys = kNumKeys * 5;
+
+  for (int k = 0; k < num_keys; ++k) {
+    int key = k + thread_id * num_keys;
+    std::string value = RandomString(&rnd, kValueSize);
+    // only the second-half is randomized
+    if (k > num_keys / 2) {
+      switch (rnd.Next() % 5) {
+        case 0:
+          write_opt.timeout_hint_us = 500 * thread_id;
+          break;
+        case 1:
+          write_opt.timeout_hint_us = num_keys - k;
+          break;
+        case 2:
+          write_opt.timeout_hint_us = 1;
+          break;
+        default:
+          write_opt.timeout_hint_us = 0;
+          state->success_kvs.insert({key, value});
+      }
+    }
+
+    uint64_t time_before_put = db->GetEnv()->NowMicros();
+    Status s = db->Put(write_opt, Key(key), value);
+    uint64_t put_duration = db->GetEnv()->NowMicros() - time_before_put;
+    if (write_opt.timeout_hint_us == 0 ||
+        put_duration + kTimerBias < write_opt.timeout_hint_us) {
+      ASSERT_OK(s);
+      std::string result;
+    }
+    if (s.IsTimedOut()) {
+      timeout_count++;
+      ASSERT_GT(put_duration + kTimerBias, write_opt.timeout_hint_us);
+    }
+  }
+
+  state->done = true;
+}
+
+TEST(DBTest, MTRandomTimeoutTest) {
+  Options options;
+  options.env = env_;
+  options.create_if_missing = true;
+  options.max_write_buffer_number = 2;
+  options.compression = kNoCompression;
+  options.level0_slowdown_writes_trigger = 10;
+  options.level0_stop_writes_trigger = 20;
+  options.write_buffer_size = kWriteBufferSize;
+  DestroyAndReopen(&options);
+
+  TimeoutWriterState thread_states[kNumThreads];
+  for (int tid = 0; tid < kNumThreads; ++tid) {
+    thread_states[tid].id = tid;
+    thread_states[tid].db = db_;
+    thread_states[tid].done = false;
+    env_->StartThread(RandomTimeoutWriter, &thread_states[tid]);
+  }
+
+  for (int tid = 0; tid < kNumThreads; ++tid) {
+    while (thread_states[tid].done == false) {
+      env_->SleepForMicroseconds(100000);
+    }
+  }
+
+  Flush();
+
+  for (int tid = 0; tid < kNumThreads; ++tid) {
+    auto& success_kvs = thread_states[tid].success_kvs;
+    for (auto it = success_kvs.begin(); it != success_kvs.end(); ++it) {
+      ASSERT_EQ(Get(Key(it->first)), it->second);
+    }
+  }
+}
+
+}  // anonymous namespace
+
 }  // namespace rocksdb
 
 int main(int argc, char** argv) {

include/rocksdb/options.h

@@ -975,7 +975,18 @@ struct WriteOptions {
   // and the write may got lost after a crash.
   bool disableWAL;
 
-  WriteOptions() : sync(false), disableWAL(false) {}
+  // If non-zero, then associated write waiting longer than the specified
+  // time MAY be aborted and returns Status::TimedOut. A write that takes
+  // less than the specified time is guaranteed to not fail with
+  // Status::TimedOut.
+  //
+  // The number of times a write call encounters a timeout is recorded in
+  // Statistics.WRITE_TIMEDOUT
+  //
+  // Default: 0
+  uint64_t timeout_hint_us;
+
+  WriteOptions() : sync(false), disableWAL(false), timeout_hint_us(0) {}
 };
 
 // Options that control flush operations

include/rocksdb/statistics.h

@@ -115,6 +115,7 @@ enum Tickers {
   // head of the writers queue.
   WRITE_DONE_BY_SELF,
   WRITE_DONE_BY_OTHER,
+  WRITE_TIMEDOUT,        // Number of writes ending up with timed-out.
   WRITE_WITH_WAL,       // Number of Write calls that request WAL
   COMPACT_READ_BYTES,   // Bytes read during compaction
   COMPACT_WRITE_BYTES,  // Bytes written during compaction
@@ -176,6 +177,7 @@ const std::vector<std::pair<Tickers, std::string>> TickersNameMap = {
     {WAL_FILE_BYTES, "rocksdb.wal.bytes"},
     {WRITE_DONE_BY_SELF, "rocksdb.write.self"},
     {WRITE_DONE_BY_OTHER, "rocksdb.write.other"},
+    {WRITE_TIMEDOUT, "rocksdb.write.timedout"},
     {WRITE_WITH_WAL, "rocksdb.write.wal"},
     {COMPACT_READ_BYTES, "rocksdb.compact.read.bytes"},
     {COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes"},

include/rocksdb/status.h

@@ -65,6 +65,12 @@ class Status {
                                    const Slice& msg2 = Slice()) {
     return Status(kShutdownInProgress, msg, msg2);
   }
+  static Status TimedOut() {
+    return Status(kTimedOut);
+  }
+  static Status TimedOut(const Slice& msg, const Slice& msg2 = Slice()) {
+    return Status(kTimedOut, msg, msg2);
+  }
 
   // Returns true iff the status indicates success.
   bool ok() const { return code() == kOk; }
@@ -93,6 +99,8 @@ class Status {
   // Returns true iff the status indicates Incomplete
   bool IsShutdownInProgress() const { return code() == kShutdownInProgress; }
 
+  bool IsTimedOut() const { return code() == kTimedOut; }
+
   // Return a string representation of this status suitable for printing.
   // Returns the string "OK" for success.
   std::string ToString() const;
@@ -106,7 +114,8 @@ class Status {
     kIOError = 5,
     kMergeInProgress = 6,
     kIncomplete = 7,
-    kShutdownInProgress = 8
+    kShutdownInProgress = 8,
+    kTimedOut = 9
   };
 
   Code code() const {

port/port_posix.cc

@@ -20,11 +20,12 @@
 namespace rocksdb {
 namespace port {
 
-static void PthreadCall(const char* label, int result) {
-  if (result != 0) {
+static int PthreadCall(const char* label, int result) {
+  if (result != 0 && result != ETIMEDOUT) {
     fprintf(stderr, "pthread %s: %s\n", label, strerror(result));
     abort();
   }
+  return result;
 }
 
 Mutex::Mutex(bool adaptive) {

util/status.cc

@@ -68,6 +68,9 @@ std::string Status::ToString() const {
     case kShutdownInProgress:
       type = "Shutdown in progress: ";
       break;
+    case kTimedOut:
+      type = "Operation timed out: ";
+      break;
     default:
       snprintf(tmp, sizeof(tmp), "Unknown code(%d): ",
                static_cast<int>(code()));

util/stop_watch.h

@@ -24,7 +24,7 @@ class StopWatch {
 
 
 
-  uint64_t ElapsedMicros() {
+  uint64_t ElapsedMicros() const {
     return env_->NowMicros() - start_time_;
   }