Add an Append call to the GCS Log that checks for current length (#1788)

* TABLE_APPEND call

* Convert callbacks back to taking in a string...

* GCS returns flatbuffers, define Log class

* Cleanups

* Modify client table to use the Log interface

* Fix bug where we replied twice from redis

* Fixes

* lint

* Compile and test raylet TaskTable

* Modify GCS tables to handle unique_ptrs from nested flatbuffers

* Add raylet::TaskTable unit tests to replace ObjectTable ones

* Convert ObjectTable to a log

* Convert ObjectTable tests to the Log

* AppendAt Redis and gcs Log command

* unit test for AppendAt

* Add a Log for task reconstruction data

* Add check for unique entries in TABLE_APPEND

* Documentation

src/ray/gcs/client.cc

@@ -17,6 +17,7 @@ Status AsyncGcsClient::Connect(const std::string &address, int port,
   object_table_.reset(new ObjectTable(context_, this));
   task_table_.reset(new TaskTable(context_, this));
   raylet_task_table_.reset(new raylet::TaskTable(context_, this));
+  task_reconstruction_log_.reset(new TaskReconstructionLog(context_, this));
   client_table_.reset(new ClientTable(context_, this, client_info));
   // TODO(swang): Call the client table's Connect() method here. To do this,
   // we need to make sure that we are attached to an event loop first. This
@@ -44,6 +45,10 @@ TaskTable &AsyncGcsClient::task_table() { return *task_table_; }
 
 raylet::TaskTable &AsyncGcsClient::raylet_task_table() { return *raylet_task_table_; }
 
+TaskReconstructionLog &AsyncGcsClient::task_reconstruction_log() {
+  return *task_reconstruction_log_;
+}
+
 ClientTable &AsyncGcsClient::client_table() { return *client_table_; }
 
 FunctionTable &AsyncGcsClient::function_table() { return *function_table_; }

src/ray/gcs/client.h

@@ -46,6 +46,7 @@ class RAY_EXPORT AsyncGcsClient {
   ObjectTable &object_table();
   TaskTable &task_table();
   raylet::TaskTable &raylet_task_table();
+  TaskReconstructionLog &task_reconstruction_log();
   ClientTable &client_table();
   inline ErrorTable &error_table();
 
@@ -65,6 +66,7 @@ class RAY_EXPORT AsyncGcsClient {
   std::unique_ptr<ObjectTable> object_table_;
   std::unique_ptr<TaskTable> task_table_;
   std::unique_ptr<raylet::TaskTable> raylet_task_table_;
+  std::unique_ptr<TaskReconstructionLog> task_reconstruction_log_;
   std::unique_ptr<ClientTable> client_table_;
   std::shared_ptr<RedisContext> context_;
   std::unique_ptr<RedisAsioClient> asio_async_client_;

src/ray/gcs/client_test.cc

@@ -215,6 +215,61 @@ TEST_F(TestGcsWithAsio, TestTableLookupFailure) {
   TestTableLookupFailure(job_id_, client_);
 }
 
+void TestLogAppendAt(const JobID &job_id, std::shared_ptr<gcs::AsyncGcsClient> client) {
+  TaskID task_id = TaskID::from_random();
+  std::vector<std::string> managers = {"A", "B"};
+  std::vector<std::shared_ptr<TaskReconstructionDataT>> data_log;
+  for (const auto &manager : managers) {
+    auto data = std::make_shared<TaskReconstructionDataT>();
+    data->node_manager_id = manager;
+    data_log.push_back(data);
+  }
+
+  // Check that we added the correct task.
+  auto failure_callback = [task_id](gcs::AsyncGcsClient *client, const UniqueID &id,
+                                    const std::shared_ptr<TaskReconstructionDataT> d) {
+    ASSERT_EQ(id, task_id);
+    test->IncrementNumCallbacks();
+  };
+
+  RAY_CHECK_OK(client->task_reconstruction_log().Append(job_id, task_id, data_log.front(),
+                                                        nullptr));
+  RAY_CHECK_OK(client->task_reconstruction_log().AppendAt(job_id, task_id, data_log[1],
+                                                          nullptr, failure_callback, 0));
+  RAY_CHECK_OK(client->task_reconstruction_log().AppendAt(job_id, task_id, data_log[1],
+                                                          nullptr, failure_callback, 2));
+  RAY_CHECK_OK(client->task_reconstruction_log().AppendAt(job_id, task_id, data_log[1],
+                                                          nullptr, failure_callback, 1));
+
+  auto lookup_callback = [task_id, managers](
+      gcs::AsyncGcsClient *client, const UniqueID &id,
+      const std::vector<TaskReconstructionDataT> &data) {
+    std::vector<std::string> appended_managers;
+    for (const auto &entry : data) {
+      appended_managers.push_back(entry.node_manager_id);
+    }
+    ASSERT_EQ(appended_managers, managers);
+    test->Stop();
+  };
+  RAY_CHECK_OK(
+      client->task_reconstruction_log().Lookup(job_id, task_id, lookup_callback));
+  // Run the event loop. The loop will only stop if the Lookup callback is
+  // called (or an assertion failure).
+  test->Start();
+  ASSERT_EQ(test->NumCallbacks(), 2);
+}
+
+TEST_F(TestGcsWithAe, TestLogAppendAt) {
+  test = this;
+  TestLogAppendAt(job_id_, client_);
+}
+
+TEST_F(TestGcsWithAsio, TestLogAppendAt) {
+  test = this;
+  TestLogAppendAt(job_id_, client_);
+}
+
+// Task table callbacks.
 void TaskAdded(gcs::AsyncGcsClient *client, const TaskID &id,
                const std::shared_ptr<TaskTableDataT> data) {
   ASSERT_EQ(data->scheduling_state, SchedulingState_SCHEDULED);

src/ray/gcs/format/gcs.fbs

@@ -10,7 +10,8 @@ enum TablePrefix:int {
   RAYLET_TASK,
   CLIENT,
   OBJECT,
-  FUNCTION
+  FUNCTION,
+  TASK_RECONSTRUCTION
 }
 
 // The channel that Add operations to the Table should be published on, if any.
@@ -35,9 +36,19 @@ table FunctionTableData {
 }
 
 table ObjectTableData {
+  // The size of the object.
   object_size: long;
+  // The node manager ID that this object appeared on or was evicted by.
   manager: string;
+  // Whether this entry is an addition or a deletion.
   is_eviction: bool;
+  // The number of times this object has been evicted from this node so far.
+  num_evictions: int;
+}
+
+table TaskReconstructionData {
+  num_executions: int;
+  node_manager_id: string;
 }
 
 enum SchedulingState:int {

src/ray/gcs/redis_context.cc

@@ -183,17 +183,29 @@ Status RedisContext::AttachToEventLoop(aeEventLoop *loop) {
 Status RedisContext::RunAsync(const std::string &command, const UniqueID &id,
                               const uint8_t *data, int64_t length,
                               const TablePrefix prefix, const TablePubsub pubsub_channel,
-                              int64_t callback_index) {
+                              int64_t callback_index, int log_length) {
   if (length > 0) {
-    std::string redis_command = command + " %d %d %b %b";
-    int status = redisAsyncCommand(
-        async_context_, reinterpret_cast<redisCallbackFn *>(&GlobalRedisCallback),
-        reinterpret_cast<void *>(callback_index), redis_command.c_str(), prefix,
-        pubsub_channel, id.data(), id.size(), data, length);
-    if (status == REDIS_ERR) {
-      return Status::RedisError(std::string(async_context_->errstr));
+    if (log_length >= 0) {
+      std::string redis_command = command + " %d %d %b %b %d";
+      int status = redisAsyncCommand(
+          async_context_, reinterpret_cast<redisCallbackFn *>(&GlobalRedisCallback),
+          reinterpret_cast<void *>(callback_index), redis_command.c_str(), prefix,
+          pubsub_channel, id.data(), id.size(), data, length, log_length);
+      if (status == REDIS_ERR) {
+        return Status::RedisError(std::string(async_context_->errstr));
+      }
+    } else {
+      std::string redis_command = command + " %d %d %b %b";
+      int status = redisAsyncCommand(
+          async_context_, reinterpret_cast<redisCallbackFn *>(&GlobalRedisCallback),
+          reinterpret_cast<void *>(callback_index), redis_command.c_str(), prefix,
+          pubsub_channel, id.data(), id.size(), data, length);
+      if (status == REDIS_ERR) {
+        return Status::RedisError(std::string(async_context_->errstr));
+      }
     }
   } else {
+    RAY_CHECK(log_length == -1);
     std::string redis_command = command + " %d %d %b";
     int status = redisAsyncCommand(
         async_context_, reinterpret_cast<redisCallbackFn *>(&GlobalRedisCallback),

src/ray/gcs/redis_context.h

@@ -53,9 +53,25 @@ class RedisContext {
   ~RedisContext();
   Status Connect(const std::string &address, int port);
   Status AttachToEventLoop(aeEventLoop *loop);
+
+  /// Run an operation on some table key.
+  ///
+  /// \param command The command to run. This must match a registered Ray Redis
+  ///        command. These are strings of the format "RAY.TABLE_*".
+  /// \param id The table key to run the operation at.
+  /// \param data The data to add to the table key, if any.
+  /// \param length The length of the data to be added, if data is provided.
+  /// \param prefix
+  /// \param pubsub_channel
+  /// \param callback_index
+  /// \param log_length The RAY.TABLE_APPEND command takes in an optional index
+  ///        at which the data must be appended. For all other commands, set to
+  ///        -1 for unused. If set, then data must be provided.
   Status RunAsync(const std::string &command, const UniqueID &id, const uint8_t *data,
                   int64_t length, const TablePrefix prefix,
-                  const TablePubsub pubsub_channel, int64_t callback_index);
+                  const TablePubsub pubsub_channel, int64_t callback_index,
+                  int log_length = -1);
+
   Status SubscribeAsync(const ClientID &client_id, const TablePubsub pubsub_channel,
                         int64_t callback_index);
   redisAsyncContext *async_context() { return async_context_; }

src/ray/gcs/tables.cc

@@ -14,6 +14,7 @@ Status Log<ID, Data>::Append(const JobID &job_id, const ID &id,
       new CallbackData({id, data, nullptr, nullptr, this, client_}));
   int64_t callback_index =
       RedisCallbackManager::instance().add([d, done](const std::string &data) {
+        RAY_CHECK(data.empty());
         if (done != nullptr) {
           (done)(d->client, d->id, d->data);
         }
@@ -26,6 +27,32 @@ Status Log<ID, Data>::Append(const JobID &job_id, const ID &id,
                             prefix_, pubsub_channel_, callback_index);
 }
 
+template <typename ID, typename Data>
+Status Log<ID, Data>::AppendAt(const JobID &job_id, const ID &id,
+                               std::shared_ptr<DataT> data, const WriteCallback &done,
+                               const WriteCallback &failure, int log_length) {
+  auto d = std::shared_ptr<CallbackData>(
+      new CallbackData({id, data, nullptr, nullptr, this, client_}));
+  int64_t callback_index =
+      RedisCallbackManager::instance().add([d, done, failure](const std::string &data) {
+        if (data.empty()) {
+          if (done != nullptr) {
+            (done)(d->client, d->id, d->data);
+          }
+        } else {
+          if (failure != nullptr) {
+            (failure)(d->client, d->id, d->data);
+          }
+        }
+        return true;
+      });
+  flatbuffers::FlatBufferBuilder fbb;
+  fbb.ForceDefaults(true);
+  fbb.Finish(Data::Pack(fbb, data.get()));
+  return context_->RunAsync("RAY.TABLE_APPEND", id, fbb.GetBufferPointer(), fbb.GetSize(),
+                            prefix_, pubsub_channel_, callback_index, log_length);
+}
+
 template <typename ID, typename Data>
 Status Log<ID, Data>::Lookup(const JobID &job_id, const ID &id, const Callback &lookup) {
   auto d = std::shared_ptr<CallbackData>(
@@ -308,6 +335,7 @@ template class Log<ObjectID, ObjectTableData>;
 template class Log<TaskID, ray::protocol::Task>;
 template class Table<TaskID, ray::protocol::Task>;
 template class Table<TaskID, TaskTableData>;
+template class Log<TaskID, TaskReconstructionData>;
 
 }  // namespace gcs
 

src/ray/gcs/tables.h

@@ -77,6 +77,23 @@ class Log {
   Status Append(const JobID &job_id, const ID &id, std::shared_ptr<DataT> data,
                 const WriteCallback &done);
 
+  /// Append a log entry to a key if and only if the log has the given number
+  /// of entries.
+  ///
+  /// \param job_id The ID of the job (= driver).
+  /// \param id The ID of the data that is added to the GCS.
+  /// \param data Data to append to the log.
+  /// \param done Callback that is called if the data was appended to the log.
+  /// \param failure Callback that is called if the data was not appended to
+  ///        the log because the log length did not match the given
+  ///        `log_length`.
+  /// \param log_length The number of entries that the log must have for the
+  ///        append to succeed.
+  /// \return Status
+  Status AppendAt(const JobID &job_id, const ID &id, std::shared_ptr<DataT> data,
+                  const WriteCallback &done, const WriteCallback &failure,
+                  int log_length);
+
   /// Lookup the log values at a key asynchronously.
   ///
   /// \param job_id The ID of the job (= driver).
@@ -241,6 +258,15 @@ using ClassTable = Table<ClassID, ClassTableData>;
 // TODO(swang): Set the pubsub channel for the actor table.
 using ActorTable = Table<ActorID, ActorTableData>;
 
+class TaskReconstructionLog : public Log<TaskID, TaskReconstructionData> {
+ public:
+  TaskReconstructionLog(const std::shared_ptr<RedisContext> &context,
+                        AsyncGcsClient *client)
+      : Log(context, client) {
+    prefix_ = TablePrefix_TASK_RECONSTRUCTION;
+  }
+};
+
 namespace raylet {
 
 class TaskTable : public Table<TaskID, ray::protocol::Task> {