Vishal Khandelwal created HBASE-18872:
-----------------------------------------
Summary: Backup scaling for multiple table and millions of row
Key: HBASE-18872
URL: https://issues.apache.org/jira/browse/HBASE-18872
Project: HBase
Issue Type: Improvement
Reporter: Vishal Khandelwal
I did a simple experiment of loading ~200 million rows on a table 1 and nothing
in a table 2. This test was done on a local cluster ~ approx 3-4 containers
were running in parallel. The focus of the test was not on how much time backup
takes but on time spent on the table were no data has been changed.
*Table without Data -->*
Elapsed: 44mins, 52sec
Average Map Time 3sec
Average Shuffle Time 2mins, 35sec
Average Merge Time 0sec
Average Reduce Time 0sec
Map : 2052
Reduce : 1
*Table with Data -->*
Elapsed: 1hrs, 44mins, 10sec
Average Map Time 4sec
Average Shuffle Time 37sec
Average Merge Time 3sec
Average Reduce Time 47sec
Map : 2052
Reduce : 64
All above numbers are a single node cluster so not many mappers run in
parallel. but let's extrapolate this to 20 node cluster, with ~100 tables and
data size to be backed up various for approx 2000 Wals, let us say each 20 node
can process 3 containers i.e 60 wals in parallel. assume 3 sec are spent in
each WALs i.e. 6000\ 60 sec --> 100 per table --> 10000 sec for all tables.
~166 hrs only for filtering. This does not seem to be scale. (These are just
rough numbers from a basic test). As all parsing is O (m (WALS) * n (Tables))
Main intend of this test is to see even the backup of very less churning table
might take good amount for just filtering the data. As number of table or data
increases, this does not seem scalable
Even i can see from our current cluster numbers easily close to 100 table, 200
millions rows, 200 -300 GB.
I would suggest that we should have filtering to parse WALs once and to
segregate in multiple WALs per table --> hFiles from per table wals. ( just a
rough idea).
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
