/// <summary>Create a new pair of key selectors using FIRST_GREATER_OR_EQUAL on both keys</summary>
public static KeySelectorPair Create(KeyRange range)
{
return(new KeySelectorPair(
KeySelector.FirstGreaterOrEqual(range.Begin),
KeySelector.FirstGreaterOrEqual(range.End)
));
}
/// <summary>Create a new pair of key selectors using FIRST_GREATER_OR_EQUAL on both keys</summary>
public static KeySelectorPair Create(Slice begin, Slice end)
{
return(new KeySelectorPair(
KeySelector.FirstGreaterOrEqual(begin),
KeySelector.FirstGreaterOrEqual(end)
));
}
protected override async ValueTask <bool> OnFirstAsync()
{
this.RemainingCount = this.Query.Limit;
this.RemainingSize = this.Query.TargetBytes;
this.Begin = this.Query.Begin;
this.End = this.Query.End;
if (this.RemainingCount == 0)
{
// we can safely optimize this case by not doing any query, because it should not have any impact on conflict resolutions.
// => The result of 'query.Take(0)' will not change even if someone adds/remove to the range
// => The result of 'query.Take(X)' where X would be computed from reads in the db, and be equal to 0, would conflict because of those reads anyway.
return(false);
}
var bounds = this.Query.OriginalRange;
// if the original range has been changed, we need to ensure that the current begin/end do not overflow:
if (this.Begin != bounds.Begin || this.End != bounds.End)
{
//TODO: find a better way to do this!
var keys = await this.Transaction.GetKeysAsync(new[] { bounds.Begin, this.Begin, bounds.End, this.End }).ConfigureAwait(false);
var min = keys[0] >= keys[1] ? keys[0] : keys[1];
var max = keys[2] <= keys[3] ? keys[2] : keys[3];
if (min >= max)
{
return(false); // range is empty
}
// rewrite the initial selectors with the bounded keys
this.Begin = KeySelector.FirstGreaterOrEqual(min);
this.End = KeySelector.FirstGreaterOrEqual(max);
}
return(true);
}
/// <summary>Create a new pair of key selectors that will select all the keys that start with the specified prefix</summary>
public static KeySelectorPair StartsWith(Slice prefix)
{
var range = KeyRange.StartsWith(prefix);
return(new KeySelectorPair(
KeySelector.FirstGreaterOrEqual(range.Begin),
KeySelector.FirstGreaterOrEqual(range.End)
));
}
/// <summary>Asynchronously fetch a new page of results</summary>
/// <returns>True if Chunk contains a new page of results. False if all results have been read.</returns>
private Task <bool> FetchNextPageAsync()
{
Contract.Requires(!this.AtEnd);
Contract.Requires(this.Iteration >= 0);
m_ct.ThrowIfCancellationRequested();
this.Transaction.EnsureCanRead();
this.Iteration++;
#if DEBUG_RANGE_PAGING
Debug.WriteLine("FdbRangeQuery.PagingIterator.FetchNextPageAsync(iter=" + this.Iteration + ") started");
#endif
var mode = this.Query.Mode;
// select the appropriate streaming mode if purpose is not default
switch (m_mode)
{
case AsyncIterationHint.Iterator:
{
// the caller is responsible for calling MoveNext(..) and deciding if it wants to continue or not..
mode = FdbStreamingMode.Iterator;
break;
}
case AsyncIterationHint.All:
{
// we are in a ToList or ForEach, we want to read everything in as few chunks as possible
mode = FdbStreamingMode.WantAll;
break;
}
case AsyncIterationHint.Head:
{
// the caller only expect one (or zero) values
mode = FdbStreamingMode.Iterator;
break;
}
}
//BUGBUG: mix the custom cancellation token with the transaction, if it is different !
var task = (this.Query.Snapshot ? this.Transaction.Snapshot : this.Transaction)
.GetRangeAsync(this.Begin, this.End, this.RemainingCount ?? 0, this.Query.Reversed, this.RemainingSize ?? 0, mode, this.Query.Read, this.Iteration)
.Then((result) =>
{
this.Chunk = result.Items;
this.RowCount += result.Count;
this.HasMore = result.HasMore;
// subtract number of row from the remaining allowed
if (this.RemainingCount.HasValue)
{
this.RemainingCount = this.RemainingCount.Value - result.Count;
}
// subtract size of rows from the remaining allowed
if (this.RemainingSize.HasValue)
{
this.RemainingSize = this.RemainingSize.Value - result.GetSize();
}
this.AtEnd = !result.HasMore || (this.RemainingCount.HasValue && this.RemainingCount.Value <= 0) || (this.RemainingSize.HasValue && this.RemainingSize.Value <= 0);
if (!this.AtEnd)
{ // update begin..end so that next call will continue from where we left...
if (this.Query.Reversed)
{
this.End = KeySelector.FirstGreaterOrEqual(result.Last);
}
else
{
this.Begin = KeySelector.FirstGreaterThan(result.Last);
}
}
#if DEBUG_RANGE_PAGING
Debug.WriteLine("FdbRangeQuery.PagingIterator.FetchNextPageAsync() returned " + this.Chunk.Length + " results (" + this.RowCount + " total) " + (hasMore ? " with more to come" : " and has no more data"));
#endif
if (!result.IsEmpty && this.Transaction != null)
{
return(Task.FromResult(Publish(result.Items)));
}
return(Completed().AsTask());
});
// keep track of this operation
this.PendingReadTask = task;
return(task);
}
/// <summary>Estimate the number of keys in the specified range.</summary>
/// <param name="db">Database used for the operation</param>
/// <param name="beginInclusive">Key defining the beginning of the range</param>
/// <param name="endExclusive">Key defining the end of the range</param>
/// <param name="onProgress">Optional callback called everytime the count is updated. The first argument is the current count, and the second argument is the last key that was found.</param>
/// <param name="cancellationToken">Token used to cancel the operation</param>
/// <returns>Number of keys k such that <paramref name="beginInclusive"/> <= k > <paramref name="endExclusive"/></returns>
/// <remarks>If the range contains a large of number keys, the operation may need more than one transaction to complete, meaning that the number will not be transactionally accurate.</remarks>
public static async Task <long> EstimateCountAsync([NotNull] IFdbDatabase db, Slice beginInclusive, Slice endExclusive, IProgress <STuple <long, Slice> > onProgress, CancellationToken cancellationToken)
{
const int INIT_WINDOW_SIZE = 1 << 8; // start at 256 //1024
const int MAX_WINDOW_SIZE = 1 << 13; // never use more than 4096
const int MIN_WINDOW_SIZE = 64; // use range reads when the windows size is smaller than 64
if (db == null)
{
throw new ArgumentNullException("db");
}
if (endExclusive < beginInclusive)
{
throw new ArgumentException("The end key cannot be less than the begin key", "endExclusive");
}
cancellationToken.ThrowIfCancellationRequested();
// To count the number of items in the range, we will scan it using a key selector with an offset equal to our window size
// > if the returned key is still inside the range, we add the window size to the counter, and start again from the current key
// > if the returned key is outside the range, we reduce the size of the window, and start again from the previous key
// > if the returned key is exactly equal to the end of range, OR if the window size was 1, then we stop
// Since we don't know in advance if the range contains 1 key or 1 Billion keys, choosing a good value for the window size is critical:
// > if it is too small and the range is very large, we will need too many sequential reads and the network latency will quickly add up
// > if it is too large and the range is small, we will spend too many times halving the window size until we get the correct value
// A few optimizations are possible:
// > we could start with a small window size, and then double its size on every full segment (up to a maximum)
// > for the last segment, we don't need to wait for a GetKey to complete before issuing the next, so we could split the segment into 4 (or more), do the GetKeyAsync() in parallel, detect the quarter that cross the boundary, and iterate again until the size is small
// > once the window size is small enough, we can switch to using GetRange to read the last segment in one shot, instead of iterating with window size 16, 8, 4, 2 and 1 (the wost case being 2^N - 1 items remaning)
// note: we make a copy of the keys because the operation could take a long time and the key's could prevent a potentially large underlying buffer from being GCed
var cursor = beginInclusive.Memoize();
var end = endExclusive.Memoize();
using (var tr = db.BeginReadOnlyTransaction(cancellationToken))
{
#if TRACE_COUNTING
tr.Annotate("Estimating number of keys in range {0}", KeyRange.Create(beginInclusive, endExclusive));
#endif
tr.SetOption(FdbTransactionOption.ReadYourWritesDisable);
// start looking for the first key in the range
cursor = await tr.Snapshot.GetKeyAsync(KeySelector.FirstGreaterOrEqual(cursor)).ConfigureAwait(false);
if (cursor >= end)
{ // the range is empty !
return(0);
}
// we already have seen one key, so add it to the count
#if TRACE_COUNTING
int iter = 1;
#endif
long counter = 1;
// start with a medium-sized window
int windowSize = INIT_WINDOW_SIZE;
bool last = false;
while (cursor < end)
{
Contract.Assert(windowSize > 0);
var selector = KeySelector.FirstGreaterOrEqual(cursor) + windowSize;
Slice next = Slice.Nil;
FdbException error = null;
try
{
next = await tr.Snapshot.GetKeyAsync(selector).ConfigureAwait(false);
#if TRACE_COUNTING
++iter;
#endif
}
catch (FdbException e)
{
error = e;
}
if (error != null)
{
// => from this point, the count returned will not be transactionally accurate
if (error.Code == FdbError.PastVersion)
{ // the transaction used up its time window
tr.Reset();
}
else
{ // check to see if we can continue...
await tr.OnErrorAsync(error.Code).ConfigureAwait(false);
}
// retry
tr.SetOption(FdbTransactionOption.ReadYourWritesDisable);
continue;
}
//BUGBUG: GetKey(...) always truncate the result to \xFF if the selected key would be past the end,
// so we need to fall back immediately to the binary search and/or get_range if next == \xFF
if (next > end)
{ // we have reached past the end, switch to binary search
last = true;
// if window size is already 1, then we have counted everything (the range.End key does not exist in the db)
if (windowSize == 1)
{
break;
}
if (windowSize <= MIN_WINDOW_SIZE)
{ // The window is small enough to switch to reading for counting (will be faster than binary search)
#if TRACE_COUNTING
tr.Annotate("Switch to reading all items (window size = {0})", windowSize);
#endif
// Count the keys by reading them. Also, we know that there can not be more than windowSize - 1 remaining
int n = await tr.Snapshot
.GetRange(
KeySelector.FirstGreaterThan(cursor), // cursor has already been counted once
KeySelector.FirstGreaterOrEqual(end),
new FdbRangeOptions()
{
Limit = windowSize - 1
}
)
.CountAsync()
.ConfigureAwait(false);
counter += n;
if (onProgress != null)
{
onProgress.Report(STuple.Create(counter, end));
}
#if TRACE_COUNTING
++iter;
#endif
break;
}
windowSize >>= 1;
continue;
}
// the range is not finished, advance the cursor
counter += windowSize;
cursor = next;
if (onProgress != null)
{
onProgress.Report(STuple.Create(counter, cursor));
}
if (!last)
{ // double the size of the window if we are not in the last segment
windowSize = Math.Min(windowSize << 1, MAX_WINDOW_SIZE);
}
}
#if TRACE_COUNTING
tr.Annotate("Found {0} keys in {1} iterations", counter, iter);
#endif
return(counter);
}
}
