/// <summary>
/// Starts reclaiming transaction log entries that are no longer reachable. Typically called in a fire-and-forget fashion.
/// </summary>
/// <returns>Task to indicate when garbage collection is finished.</returns>
public async Task ReclaimAsync()
{
_parent.AssertInvariants();
if (_parent._heldCount > _parent._activeCount)
{
using (await _parent._lock.EnterAsync().ConfigureAwait(false))
{
if (_parent._heldCount > _parent._activeCount)
{
Tracing.Transaction_Log_Garbage_Collection_Start(null, _parent._engineId, _parent._latest, _parent._activeCount, _parent._heldCount);
var difference = _parent._heldCount - _parent._activeCount;
using (var tx = _parent._keyValueStore.CreateTransaction())
{
foreach (var table in _parent._versionedLogs.Take((int)difference))
{
table.Scope(tx).Clear();
}
var metadata = _parent._metadataTable.Enter(tx);
metadata.Update(HeldCountKey, _parent._activeCount.ToString(CultureInfo.InvariantCulture));
await tx.CommitAsync().ConfigureAwait(false);
Tracing.Transaction_Log_Garbage_Collection(null, _parent._engineId, difference, _parent._latest, _parent._activeCount);
for (var i = 0; i < difference; i++)
{
_parent._versionedLogs.Dequeue();
}
_parent._heldCount = _parent._activeCount;
}
Tracing.Transaction_Log_Garbage_Collection_End(null, _parent._engineId, _parent._latest, _parent._activeCount, _parent._heldCount);
}
}
}
_parent.AssertInvariants();
}
/// <summary>
/// Triggers truncation of the transaction log (everything up to and including the current snapshot that has been
/// superseded by the state persisted by the successful checkpoint).
/// </summary>
/// <param name="transaction">The transaction to use to update the underlying tables.</param>
/// <returns>
/// Task completing when metadata has been updated to indicate the intent to truncate; the returned object is then used
/// to trigger a background GC using <see cref="ReclaimResource.ReclaimAsync"/>.
/// </returns>
/// <remarks>
/// If a failure happens after a successful checkpoint but before calling (and completing) the lose reference operation,
/// subsequent recovery will cause coalescing of the not-yet-discarded snapshotted-but-checkpointed log entries with any
/// newer entries that were added. We're hardended against these operations that will cause "create existing" and "delete
/// non-existing" resources, but it's not ideal.
///
/// When/if we supersede the IStateWriter approach to checkpoints and fully commit to using the key/value store passed to
/// the engine's constructor, we can have a transaction that spans both updates, and not have to worry about this edge case.
///
/// The reason for having the stores in two places is historical when early versions of Reaqtor managed the transaction log
/// externally to an engine using specialized transaction log facilities (akin to CLFS) that were not integrated with the
/// key/value store used (i.e. Service Fabric KVS). Over time, we moved away from this model, but we had to keep the
/// IStateWriter and IStateReader that's used in other environments as well.
///
/// It's worth considering an alternative engine implementation (alongside the current one for starters) that's "active"
/// and doesn't have to be told when to checkpoint either. It simply uses the key/value store passed to its constructor,
/// supports `RecoverAsync()` to recover from the given store (or has a static factory to do so, rather than a separate
/// construction step), and performs checkpointing when it needs to, akin to a GC deciding it's time to perform a GC. It
/// can do this by measuring the amount of dirty state and maybe have a configure maximum delay in between checkpoints to
/// ensure recovery time is bounded (and replay of events is limited). It can furthermore have probes on ingress pieces
/// (reachable through some IIngressProbe interface implemented by operators, so it can hunt for these using visitors over
/// the subscriptions and subjects) to compute metrics and be self-tuning. There can still be a `CheckpointAsync` which
/// is similar to `GC.Collect()` where an external party can trigger a checkpoint, e.g. right before unloading the engine
/// or when a known burst of events has passed through, so it's worth evacuating state to disk.
/// </remarks>
public async Task <ReclaimResource> LoseReferenceAsync(IKeyValueStoreTransaction transaction = null)
{
_parent.AssertInvariants();
var createdNewTransaction = false;
if (transaction == null)
{
createdNewTransaction = true;
transaction = _parent._keyValueStore.CreateTransaction();
}
// Very important to make the active count = 1 at the very least. Garbage collection is secondary (but important too).
using (await _parent._lock.EnterAsync().ConfigureAwait(false))
{
var tx = _parent._metadataTable.Enter(transaction);
try
{
tx.Update(ActiveCountKey, "1");
await transaction.CommitAsync().ConfigureAwait(false);
}
finally
{
if (createdNewTransaction)
{
transaction.Dispose();
}
}
_parent._activeCount = 1;
}
Tracing.Transaction_Log_Lost_Reference(null, _parent._engineId, _parent._latest, _parent._activeCount, _parent._heldCount);
_parent.AssertInvariants();
return(new ReclaimResource(_parent));
}
