flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Oliver Swoboda <oswobod...@gmail.com>
Subject Using Flink with Accumulo
Date Thu, 03 Nov 2016 12:57:35 GMT

I'm using Flink with Accumulo and wanted to read data from the database by
using the createHadoopInput function. Therefore I configure an
AccumuloInputFormat. The source code you can find here: https://github.com/

I'm using a 5 Node Cluster (1 Master, 4 Worker).
Accumulo is installed with Ambari and has 1 Master Server on the Master
Node and 4 Tablet Servers (one on each Worker).
Flink is installed standalone with the Jobmanager on the Master Node and 4
Taskmanagers (one on each Worker). Every Taskmanager can have 4 Tasks, so
there are 32 in total.

First problem I have:
If I start serveral Flink Jobs the client count for Zookeeper in the
Accumulo Overview is constantly increasing. I assume that the used scanner
isn't correctly closed. The client count only decreases to normal values
when I restart Flink.

Second problem I have:
I want to compare aggregations on time series data with Accumulo (with
Iterators) and with flink. Unfortunately, the results vary inexplicable
when I'm using Flink. I wanted to compare the results for a full table scan
(called baseline in the code), but sometimes it takes 17-18 minutes and
sometimes its between 30 and 60 minutes. In the longer case I can see in
the Accumulo Overview that after some time only one worker is left with
running scans and there are just a few entries/s sanned (4 million at the
beginning when all workers are running to 200k when the one worker is
left). Because there are 2.5 billion records to scan and almost 500 million
left it takes really long.
This problem doesn't occur with Accumulo using Iterators and a batch
scanner on the master node, each scan has almost identical durations and
graphics in the Accumulo Overview for entries/s, MB/s scanned and seeks are
for each scan the same.

Yours faithfully,
Oliver Swoboda

View raw message