public void Sampler_Basics()
{
Random r = new Random(8);
ArraySelector all = ArraySelector.All(10240);
int[] eighthArray = null;
ArraySelector eighth = Sampler.Eighth(all, r, ref eighthArray);
AssertClose(all.Count / 8, eighth.Count, 0.2f);
int[] sixtyfourthArray = null;
ArraySelector sixtyfourth = Sampler.Eighth(eighth, r, ref sixtyfourthArray);
AssertClose(eighth.Count / 8, sixtyfourth.Count, 0.2f);
}
public void Sample()
{
Random r = new Random(8);
ArraySelector all = ArraySelector.All(10240);
int[] eighthArray = null;
int[] sixtyfourthArray = null;
using (Benchmarker b = new Benchmarker($"Sampler.Eighth", DefaultMeasureMilliseconds))
{
b.Measure("Sampler.Eighth", all.Count, () =>
{
ArraySelector eighth = Sampler.Eighth(all, r, ref eighthArray);
ArraySelector sixtyfourth = Sampler.Eighth(eighth, r, ref sixtyfourthArray);
return(sixtyfourth.Count);
});
}
}
/// <summary>
/// Build a GroupBy Dictionary for Peek.
/// </summary>
/// <remarks>
/// Peek identifies each distinct common value and the approximate percentage of rows with it.
/// If we have many matching rows, we can sample - the sample will have any common values in it.
/// However, we don't know how many matches we have in advance.
/// Therefore, we build a Dictionary of all rows, 1/8 of rows, 1/64 of rows, and 1/512 of rows.
/// As soon as a given sample has enough samples to be statistically valid, we stop collecting the larger subsets.
/// This strategy allows us to run the overall query only once, end up with a large enough sample, and avoid building giant Dictionaries.
/// </remarks>
/// <param name="cancellationToken">CancellationToken to request early stop</param>
private void BuildDictionary(CancellationToken cancellationToken)
{
// Short-circuit path if there's one key column and it's an EnumColumn
if (_column.IsEnumColumn())
{
BuildSingleEnumColumnDictionary(cancellationToken);
return;
}
// Build a Random instance to sample rows
Random r = new Random();
// Build a Dictionary and CountAggregator for each sample
GroupByDictionary[] dictionaries = new GroupByDictionary[SampleCount];
CountAggregator[] counts = new CountAggregator[SampleCount];
int[][] remapArrays = new int[SampleCount][];
for (int i = 0; i < SampleCount; ++i)
{
dictionaries[i] = new GroupByDictionary(new ColumnDetails[] { _column.ColumnDetails });
counts[i] = new CountAggregator();
}
// Retrieve the column getter
Func <XArray> columnGetter = _column.CurrentGetter();
// Track which sample we'll currently report
int currentSample = 0;
XArray[] arrays = new XArray[1];
int count;
while ((count = _source.Next(XTableExtensions.DefaultBatchSize, cancellationToken)) != 0)
{
// Get the column values
arrays[0] = columnGetter();
// Build the GroupBy count for all rows and successive 1/8 samples
for (int i = 0; i < SampleCount; ++i)
{
// Add these to the Join Dictionary
if (i >= currentSample)
{
// Choose buckets for each row
XArray indicesForRows = dictionaries[i].FindOrAdd(arrays);
// Identify the bucket for each row and aggregate them
counts[i].Add(indicesForRows, dictionaries[i].Count);
// If this sample now has enough values, stop collecting bigger row sets
if (currentSample == i - 1 && counts[i].TotalRowCount > RequiredSampleSize)
{
// If every row was unique, stop early and don't set outputs (zero rows)
if (ShouldStopEarly(dictionaries[currentSample], counts[currentSample]))
{
return;
}
dictionaries[currentSample] = null;
counts[currentSample] = null;
currentSample++;
}
}
// Each successive dictionary has ~1/8 of the rows of the previous one
if (i < SampleCount - 1)
{
ArraySelector sample = Sampler.Eighth(arrays[0].Selector, r, ref remapArrays[i]);
arrays[0] = arrays[0].Reselect(sample);
}
}
}
// Once the loop is done, get the distinct values and aggregation results
PostSortAndFilter(dictionaries[currentSample].DistinctKeys()[0], counts[currentSample].Values, counts[currentSample].TotalRowCount, currentSample == 0);
}
