drill-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Andries Engelbrecht <aengelbre...@maprtech.com>
Subject Re: Documentation for Query Profile page in Web UI
Date Tue, 21 Apr 2015 22:08:34 GMT

Definitely looks like the majority of time is by far spend on reading the Hive data (Hive_Sub_Scan).
Not sure how well the storage environment is configured, and it may very likely be that the
nodes are just waiting on storage IO. The more nodes will simply just wait longer to actually
get the data through the same pipe. Is the data on S3 with an Hive external table or similar?

Very short answers, probably will need much more detailed documentation to cover these in
better detail than a mailing list
i) Probably will need to get some documentation to clarify that better. 
ii) Typically seconds are x.xxx and minutes are x:xx
iii) Graphical representation of fragments and time taken in seconds
iv) Best way for me to describe is that Major Fragment is a query phase (operator), and minor
fragments are the parallelization of Major fragments, that is why some Major Frags only have
a single minor frag as it can’t or doesn’t make sense to parallelize.
v) Since you have parallel execution for Major Frags it shows when the first execution thread
started, when the last one started, and similar for ends. Basically just to give you an indication
if one of more threads were significantly slower than others (important to understand for
some environments to spot a lagger that holds the whole process up).
vi) Same logic as v to see the avg execution of all threads and the quickest and slowest executors
vii) The mem is to indicate how much was assigned to a thread to complete the operation. A
simple scan operation doesn’t require that much memory vs joins and other functions than
can require much more memory


On Apr 21, 2015, at 2:34 PM, Alexander Zarei <alexanderz.simba@gmail.com> wrote:

> Hi Team Drill!
> While performing performance testing on Drill clusters on AWS EMR, with TPC-H data of
scale factor 100, I observed the results for a cluster of 3 nodes are similar to a cluster
of 13 nodes. Hence, I am investigating how the query is being carried out and which part of
the query handling (e.g. planning, reading the data, executing the query, transferring the
record batches) is the dominant time consuming part.
> Parth kindly suggested I should use the Query Profile from the Web UI and it helped a
lot. However, there are some items on the Query Profile page that I did not find documentation
to interpret them. I was wondering if you know what the following item are:
> I)                    What are the meaning of operator types: Project, Unordered receiver,
Single Sender? I guess Hive sub Scan is the time spent reading data from Hive, is that correct?
> II)                  What are the units for the Processes columns in the Operator Profiles
Overview table? Is it time in a minutes : seconds format?
> Also it would be really nice to know:
> III)                What metric does the blue chart on the top of the Overview section
> IV)               What is fragment, a minor fragment and major fragment?
> V)                 What are the first start and last start and first end and last end?
> VI)               What are the sets over which max, min and average are calculated?
> VII)             Why the Peak memory is so small? 4MB while the machine has 16 GB of
> The print of the Web UI as well as the json profile are attached.
> Thanks a lot for your time and help.
> Thanks,
> Alex
> <Apache Drill Query Profile.pdf><query profile.json>

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message