private async Task ReloadTrainingDataHighAvailabilityAsync(bool finalize, CancellationToken cancellationToken)
{
using (var trainingDataUpgradeableLock = await m_trainingDataLockAsync.EnterUpgradeableReadLockAsync(cancellationToken).ConfigureAwait(false))
{
Logging.Log.Info("Reloading training data and retraining rec sources. Rec sources will remain available.");
Logging.Log.InfoFormat("Memory use: {0} bytes", GC.GetTotalMemory(forceFullCollection: false));
Stopwatch totalTimer = Stopwatch.StartNew();
// Load new training data
MalTrainingData newData;
IDictionary <int, string> newUsernames;
IDictionary <int, IList <int> > newPrereqs;
using (IMalTrainingDataLoader malTrainingDataLoader = m_trainingDataLoaderFactory.GetTrainingDataLoader())
using (CancellationTokenSource faultCanceler = new CancellationTokenSource())
using (CancellationTokenSource faultOrUserCancel = CancellationTokenSource.CreateLinkedTokenSource(cancellationToken, faultCanceler.Token))
{
Stopwatch trainingDataTimer = Stopwatch.StartNew();
CancellableTask <MalTrainingData> trainingDataTask = new CancellableTask <MalTrainingData>(
malTrainingDataLoader.LoadMalTrainingDataAsync(faultOrUserCancel.Token), faultCanceler);
Task trainingDataTimerTask = trainingDataTask.Task.ContinueWith(task =>
{
trainingDataTimer.Stop();
Logging.Log.InfoFormat("Training data loaded. {0} users, {1} animes, {2} entries. Took {3}.",
task.Result.Users.Count, task.Result.Animes.Count,
task.Result.Users.Keys.Sum(userId => task.Result.Users[userId].Entries.Count),
trainingDataTimer.Elapsed);
},
cancellationToken, TaskContinuationOptions.ExecuteSynchronously | TaskContinuationOptions.NotOnCanceled
| TaskContinuationOptions.NotOnFaulted, TaskScheduler.Current);
Stopwatch prereqsTimer = Stopwatch.StartNew();
CancellableTask <IDictionary <int, IList <int> > > prereqsTask = new CancellableTask <IDictionary <int, IList <int> > >(
malTrainingDataLoader.LoadPrerequisitesAsync(faultOrUserCancel.Token), faultCanceler);
Task prereqsTimerTask = prereqsTask.Task.ContinueWith(task =>
{
prereqsTimer.Stop();
int numPrereqs = task.Result.Values.Sum(prereqList => prereqList.Count);
Logging.Log.InfoFormat("Prerequisites loaded. {0} prerequisites for {1} animes. Took {2}.",
numPrereqs, task.Result.Count, prereqsTimer.Elapsed);
},
cancellationToken, TaskContinuationOptions.ExecuteSynchronously | TaskContinuationOptions.NotOnCanceled
| TaskContinuationOptions.NotOnFaulted, TaskScheduler.Current);
await AsyncUtils.WhenAllCancelOnFirstExceptionDontWaitForCancellations(trainingDataTask, prereqsTask);
newData = trainingDataTask.Task.Result;
newUsernames = GetUsernamesFromTrainingData(newData);
newPrereqs = prereqsTask.Task.Result;
await trainingDataTimerTask.ConfigureAwait(false);
await prereqsTimerTask.ConfigureAwait(false);
}
GC.Collect();
Logging.Log.InfoFormat("Memory use: {0} bytes", GC.GetTotalMemory(forceFullCollection: false));
using (var recSourcesUpgradeableLock = await m_recSourcesLockAsync.EnterUpgradeableReadLockAsync(cancellationToken).ConfigureAwait(false))
{
// clone the json rec sources without the training state and train each one with the new data.
Dictionary <string, ITrainableJsonRecSource> newRecSources = new Dictionary <string, ITrainableJsonRecSource>(StringComparer.OrdinalIgnoreCase);
Dictionary <string, Func <ITrainableJsonRecSource> > newRecSourceFactories = new Dictionary <string, Func <ITrainableJsonRecSource> >(m_recSourceFactories, StringComparer.OrdinalIgnoreCase);
if (m_recSourceFactories.Count == 0)
{
Logging.Log.Info("No rec sources to retrain.");
}
else
{
Logging.Log.Info("Retraining rec sources.");
object newRecSourcesLockAndMemFence = new object();
List <Task> recSourceTrainTasksList = new List <Task>();
// ToList() so we can unload a rec source as we iterate if it errors while training.
foreach (string recSourceNameLoopVar in m_recSourceFactories.Keys.ToList())
{
string recSourceName = recSourceNameLoopVar; // avoid capturing the loop var
ITrainableJsonRecSource recSource = newRecSourceFactories[recSourceName]();
Task recSourceTrainTask = Task.Run(() =>
{
Logging.Log.InfoFormat("Retraining rec source {0} ({1}).", recSourceName, recSource);
Stopwatch trainTimer = Stopwatch.StartNew();
try
{
recSource.Train(newData, newUsernames, cancellationToken);
trainTimer.Stop();
Logging.Log.InfoFormat("Trained rec source {0} ({1}). Took {2}.", recSourceName, recSource, trainTimer.Elapsed);
lock (newRecSourcesLockAndMemFence)
{
newRecSources[recSourceName] = recSource;
}
}
catch (OperationCanceledException)
{
Logging.Log.InfoFormat("Canceled while retraining rec source {0} ({1}).", recSourceName, recSource);
throw;
}
catch (Exception ex)
{
Logging.Log.ErrorFormat("Error retraining rec source {0} ({1}): {2} Unloading it.",
ex, recSourceName, recSource, ex.Message);
lock (newRecSourcesLockAndMemFence)
{
newRecSourceFactories.Remove(recSourceName);
}
}
}, cancellationToken);
recSourceTrainTasksList.Add(recSourceTrainTask);
}
// Wait for all to complete or cancellation. There should not be any exceptions other than OperationCanceledException.
await Task.WhenAll(recSourceTrainTasksList);
lock (newRecSourcesLockAndMemFence)
{
; // just for the fence
}
}
// Swap in the newly trained rec sources.
using (var trainingDataWriteLock = await m_trainingDataLockAsync.UpgradeToWriteLock(cancellationToken).ConfigureAwait(false))
using (var recSourcesWriteLock = await m_recSourcesLockAsync.UpgradeToWriteLock(cancellationToken).ConfigureAwait(false))
{
m_recSources = newRecSources;
m_recSourceFactories = newRecSourceFactories;
m_animes = newData.Animes;
m_prereqs = newPrereqs;
if (finalize)
{
m_trainingData = null;
m_usernames = null;
m_finalized = true;
Logging.Log.Info("Finalized rec sources.");
}
else
{
m_trainingData = newData;
m_usernames = newUsernames;
m_finalized = false;
}
}
}
totalTimer.Stop();
Logging.Log.InfoFormat("All rec sources retrained with the latest data. Total time: {0}", totalTimer.Elapsed);
}
GC.Collect();
Logging.Log.InfoFormat("Memory use: {0} bytes", GC.GetTotalMemory(forceFullCollection: false));
}
public async Task LoadRecSourceAsync(Func <ITrainableJsonRecSource> recSourceFactory, string name, bool replaceExisting, CancellationToken cancellationToken)
{
// Acquire read lock on current list, write lock on pending list
// If name already exists on current list and replaceExisting = false, throw.
// If name already exists on pending list, throw.
// Otherwise, add name to pending list, release locks, and proceed.
using (var recSourcesReadLock = await m_recSourcesLockAsync.EnterReadLockAsync(cancellationToken).ConfigureAwait(false))
using (var pendingRecSourcesWriteLock = await m_pendingRecSourcesLockAsync.EnterWriteLockAsync(cancellationToken).ConfigureAwait(false))
{
if (m_recSources.ContainsKey(name) && !replaceExisting)
{
throw new RecServiceErrorException(new Error(errorCode: ErrorCodes.Unknown,
message: string.Format("A recommendation source with the name \"{0}\" already exists.", name)));
}
if (m_pendingRecSources.Contains(name))
{
throw new RecServiceErrorException(new Error(errorCode: ErrorCodes.Unknown,
message: string.Format("A recommendation source with the name \"{0}\" is currently being trained.", name)));
}
m_pendingRecSources.Add(name);
}
try
{
// Need to hold read lock on training data while training so that a retrain can't happen while we're training here.
// Rec sources must be trained with the current m_trainingData, not an old version.
using (var trainingDataReadLock = await m_trainingDataLockAsync.EnterReadLockAsync(cancellationToken).ConfigureAwait(false))
{
if (m_trainingData == null && !m_finalized)
{
throw new RecServiceErrorException(new Error(errorCode: ErrorCodes.NoTrainingData,
message: "A reload/retrain in low memory mode failed, leaving the rec service without training data or rec sources. Issue a ReloadTrainingData command to load training data, then load rec sources."));
}
else if (m_trainingData == null && m_finalized)
{
throw new RecServiceErrorException(new Error(errorCode: ErrorCodes.Finalized,
message: "Rec sources have been finalized. A non-finalized retrain must be invoked to be able to add rec sources."));
}
ITrainableJsonRecSource recSource = recSourceFactory();
Logging.Log.InfoFormat("Training rec source with name \"{0}\", replaceExisting={1}: {2}", name, replaceExisting, recSource);
Stopwatch timer = Stopwatch.StartNew();
recSource.Train(m_trainingData, m_usernames, cancellationToken);
timer.Stop();
Logging.Log.InfoFormat("Trained rec source {0}. Took {1}", name, timer.Elapsed);
using (var recSourcesWriteLock = await m_recSourcesLockAsync.EnterWriteLockAsync(cancellationToken).ConfigureAwait(false))
using (var pendingRecSourcesWriteLock = await m_pendingRecSourcesLockAsync.EnterWriteLockAsync(cancellationToken).ConfigureAwait(false))
{
m_recSources[name] = recSource;
m_recSourceFactories[name] = recSourceFactory;
m_pendingRecSources.Remove(name);
}
}
}
catch (Exception)
{
m_pendingRecSources.Remove(name);
throw;
}
Logging.Log.InfoFormat("Loaded rec source {0}.", name);
GC.Collect();
Logging.Log.InfoFormat("Memory use: {0} bytes", GC.GetTotalMemory(forceFullCollection: false));
}