Bin Shi created PHOENIX-4912:
--------------------------------
Summary: Make Table Sampling algorithm to accommodate to the imbalance row distribution across guide posts
Key: PHOENIX-4912
URL: https://issues.apache.org/jira/browse/PHOENIX-4912
Project: Phoenix
Issue Type: Improvement
Affects Versions: 5.0.0
Reporter: Bin Shi
Assignee: Bin Shi
The current implementation of table sampling is based on the assumption "Every two consecutive guide posts contains the equal number of rows" which isn't accurate in practice, and once we collect multiple versions of cells and the deleted rows, the thing will become worse.
In details, the current implementation of table sampling is (see BaseResultIterators.getParallelScan() which calls sampleScans(...) at the end of function) as described below:
# Iterate all parallel scans generated;
# For each scan, if getHashHode(start row key of the scan) MOD 100 < tableSamplingRate (See TableSamplerPredicate.java) then pick this scan; otherwise discard this scan.
The problem can be formalized as: We have a group of scans and each scan is defined as . Now we want to randomly pick X groups so that the sum of count of rows in the selected groups is close to Y, where Y = the total count of rows of all scans T * table sampling rate R.
To resolve the above problem, one of algorithms that we can consider are described below:
ArrayList TableSampling(ArrayList scans, T, R)
{
ArrayList pickedScans = new ArrayList();
Y = T * R;
for (scan in scans) {
if (Y <= 0) break;
if (getHashCode(Ki) MOD 100 < R) {
// then pick this scan, and adjust T, R, Y accordingly
pickedScans.Add(scan);
T -= Ci;
Y -= Ci;
if (T != 0 && Y > 0) {
R = Y / T;
}
}
}
return pickedScans;
}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)