Dan Burkert created KUDU-2253:
---------------------------------
Summary: Deltafile on-disk size is 3x larger than expected for
large-value workloads
Key: KUDU-2253
URL: https://issues.apache.org/jira/browse/KUDU-2253
Project: Kudu
Issue Type: Bug
Components: tablet
Affects Versions: 1.6.0
Reporter: Dan Burkert
While looking into the performance of the integration test written for
KUDU-2251 (https://gerrit.cloudera.org/#/c/8951/ revision 6), Todd and I found
that the on-disk deltafiles written are about 3x larger than expected. The
culprit is an optimization in the CFile value index which is [turned off for
delta
files|https://github.com/apache/kudu/blob/branch-1.6.x/src/kudu/tablet/deltafile.cc#L106-L107].
The optimization truncates large keys after the first unique byte between
sequential values. The deltafile values, in the case of this integration test,
include the small {{DeltaKey}}, and the 8KiB updated value. As a result the
BTree interior nodes are being completely filled by only ~4 values ([32KiB
cblock size by
default|https://github.com/apache/kudu/blob/branch-1.6.x/src/kudu/tablet/deltafile.cc#L58-L60]).
This makes the BTree far less effective, and means that the full updated data
is written many times. We expect fixing this will improve performance for
update-heavy workloads with large values (for example YCSB).
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
