Subject: HBase Scan consumes high cpu


Are you keeping the deleted cells? Check 'VERSIONS' for the column family and set it to 1 if you don't want to keep the deleted cells.

From: [EMAIL PROTECTED] At: 09/12/19 12:40:01To:  [EMAIL PROTECTED]
Subject: Re: HBase Scan consumes high cpu

Hi,

As said earlier, we have populated the rowkey "MY_ROW" with integers
from 0 to 1500000 as column qualifiers. Then we have deleted the
qualifiers from 0 to 1499000.

We executed the following query. It took 15.3750 seconds to execute.

hbase(main):057:0> get 'mytable', 'MY_ROW', {COLUMN=>['pcf'],
FILTER=>ColumnRangeFilter.new(Bytes.toBytes(1499000.to_java(:int)),
true, Bytes.toBytes(1499010.to_java(:int)), false)}
COLUMN                            CELL
  pcf:\x00\x16\xDFx                timestamp=1568123881899,
value=\x00\x16\xDFx
  pcf:\x00\x16\xDFy                timestamp=1568123881899,
value=\x00\x16\xDFy
  pcf:\x00\x16\xDFz                timestamp=1568123881899,
value=\x00\x16\xDFz
  pcf:\x00\x16\xDF{                timestamp=1568123881899,
value=\x00\x16\xDF{
  pcf:\x00\x16\xDF|                timestamp=1568123881899,
value=\x00\x16\xDF|
  pcf:\x00\x16\xDF}                timestamp=1568123881899,
value=\x00\x16\xDF}
  pcf:\x00\x16\xDF~                timestamp=1568123881899,
value=\x00\x16\xDF~
  pcf:\x00\x16\xDF\x7F             timestamp=1568123881899,
value=\x00\x16\xDF\x7F
  pcf:\x00\x16\xDF\x80             timestamp=1568123881899,
value=\x00\x16\xDF\x80
  pcf:\x00\x16\xDF\x81             timestamp=1568123881899,
value=\x00\x16\xDF\x81
1 row(s) in 15.3750 seconds
Now we inserted a new column with qualifier 10 (\x0A), such that it
comes earlier in lexicographical order. Now we executed the same query.
It only took 0.0240 seconds.

hbase(main):058:0> put 'mytable', 'MY_ROW', "pcf:\x0A", "\x00"
0 row(s) in 0.0150 seconds
hbase(main):059:0> get 'mytable', 'MY_ROW', {COLUMN=>['pcf'],
FILTER=>ColumnRangeFilter.new(Bytes.toBytes(1499000.to_java(:int)),
true, Bytes.toBytes(1499010.to_java(:int)), false)}
COLUMN                            CELL
  pcf:\x00\x16\xDFx                timestamp=1568123881899,
value=\x00\x16\xDFx
  pcf:\x00\x16\xDFy                timestamp=1568123881899,
value=\x00\x16\xDFy
  pcf:\x00\x16\xDFz                timestamp=1568123881899,
value=\x00\x16\xDFz
  pcf:\x00\x16\xDF{                timestamp=1568123881899,
value=\x00\x16\xDF{
  pcf:\x00\x16\xDF|                timestamp=1568123881899,
value=\x00\x16\xDF|
  pcf:\x00\x16\xDF}                timestamp=1568123881899,
value=\x00\x16\xDF}
  pcf:\x00\x16\xDF~                timestamp=1568123881899,
value=\x00\x16\xDF~
  pcf:\x00\x16\xDF\x7F             timestamp=1568123881899,
value=\x00\x16\xDF\x7F
  pcf:\x00\x16\xDF\x80             timestamp=1568123881899,
value=\x00\x16\xDF\x80
  pcf:\x00\x16\xDF\x81             timestamp=1568123881899,
value=\x00\x16\xDF\x81
1 row(s) in 0.0240 seconds
hbase(main):060:0>
We were able to reproduce the result consistently same, the pattern
being bulk insert followed by bulk delete of most of the earlier columns.
We observed the following behaviour while debugging the StoreScanner
(regionserver).

Case 1:

1. When StoreScanner.next() is called, it starts to iterate over the
cells from the start of the rowkey.

2. As all the cells are deleted (from 0 to 1499000), we could see
alternate delete and put type cells. Now, the
NormalUserScanQueryMatcher.match() returns
ScanQueryMatcher.MatchCode.SKIP and
ScanQueryMatcher.MatchCode.SEEK_NEXT_COL for Delete and Put type cell
respectively. This iteration happens throughout the range of 0 to 1499000.

3. This happens until a valid Put type cell is encountered, where the
matcher applies the ColumnRangeFilter to the cell, which in turm returns
ScanQueryMatcher.MatchCode.SEEK_NEXT_USING_HINT. In the next iteration
it seeks directly to the desired column.
Case 2:

1. When StoreScanner.next() is called, it starts to iterate over the
cells from the start of the rowkey.

2. When the Put cell of qualifier 10 (\x0A) is encountered, the matcher
returns ScanQueryMatcher.MatchCode.SEEK_NEXT_USING_HINT. In the next
iteration it seeks directly to the desired column.
Please let us know if this behaviour is intentional or it could be avoided.

Regards,

Solvannan R M
On 2019/09/10 17:12:36, Josh Elser wrote:
 > Deletes are held in memory. They represent data you have to traverse >
 > until that data is flushed out to disk. When you write a new cell
with a >
 > qualifier of 10, that sorts, lexicographically, "early" with respect
to >
 > the other qualifiers you've written.>
 >
 > By that measure, if you are only scanning for the first column in this >
 > row which you've loaded with deletes, it would make total sense to me >
 > that the first case is slow and the second fast is fast>
 >
 > Can you please share exactly how you execute your "query" for
both(all) >
 > scenarios?>
 >
 > On 9/10/19 11:35 AM, Solvannan R M wrote:>
 > > Hi,>
 > > >
 > > We have been using HBase (1.4.9) for a case where timeseries data
is continuously inserted and deleted (high churn) against a single
rowkey. The column keys would represent timestamp more or less. When we
scan this data using ColumnRangeFilter for a recent time-range, scanner
for the stores (memstore & storefiles) has to go through contiguous
deletes, before it reaches the requested timerange data. While using
this scan, we could notice 100% cpu usages in single core by the
regionserver process.>
 > > >
 > > So, for our case, most of the cells with older timestamps will be
in deleted state. While traversing these deleted cells, the regionserver
process causing 100% cpu usage in single core.>
 > > >
 > > We tried to trace the code for scan and we observed the following
behaviour.>
 > > >
 > > 1. While scanner is initialized, it seeked all the store-scanners
to the start of the rowkey.>
 > > 2. Then it traverses the deleted cells and discards it (as it was
deleted) one by one.>
 > > 3. When it