Hi Tilak,
Since we only had your 10 TB and 5 minute numbers to play with, I did the
back-of-envelope calcs using those.
Of course, you can/should look for ways to reduce the data read per query.
There are two places to start: vertical and horizontal partitioning.
Vertical partitioning comes from storing the data in the columnar Parquet
format. Suppose your data has 100 columns, but a typical query uses some
combination of just five columns. Parquet will read just those columns, reading
just 5/100 = 5% of the data or 500 GB. Assuming good distribution of data
blocks on HDFS, that greatly reduces the number of spindles needed to meet
SLAs: down from ~300 to ~15.
Horizontal partitioning comes from doing directory partitioning. The most
common pattern is by date for time series data. Suppose your 10 TB represents
three years of data. If data is partitioned by date, you'll get roughly 1000
partitions. If each query touches just one day of data, you'll read just 0.1%
of the data. If a query touches a month, you'll read just 3% of data.
The good news is that the two forms of partitioning combine. If you apply both
forms of partitioning to our hypothetical example, you reduce the scan to 3%
(for the scan over one month of data) * 5% (for reading 5/100 columns) = 0.15%
of data or just 15 GB of data. This can be done in far less than five minutes.
Of course, your results completely depend on your data and queries. And, at
this level of partitioning, other costs begin to dominate such as the metadata
preparation discussed in a previous message, the cost of doing sorts,
aggregations, etc. and so on. Good measurements will tell you the now-dominant
costs.
Thanks,
- Paul
On Monday, August 13, 2018, 1:17:32 AM PDT, Surneni Tilak
<[email protected]> wrote:
Hi Paul,
Thanks a lot for your response which I feel Great as I am looking for this
kind of approach in specific to reduce my query execution time.
According to your response I could have a check on multiple hardware related
parameters that are being used in my cluster. It's true that I did not provide
much of details regarding my cluster but thanks for understanding my query. I
will try out by tuning these parameters and will reach out to our group if I
need any further help.
Best regards,
__________
Tilak
-----Original Message-----
From: Paul Rogers [mailto:[email protected]]
Sent: Monday, August 13, 2018 2:20 AM
To: [email protected]
Subject: RE: Drill Configuration Requirements To Query Data in Tera Bytes
Hi Tilak,
Quick follow-up to this old thread. Few details were provided other than a
desire to scan 10 TB of data in 5 minutes. Here is a quick back of the envelope
calculation.
Assume that the Drill query is the only operation active on your cluster.
Assume you are doing a simple SELECT query with no joins or sorts. You must get
the entire 10 TB of data off disk and into Drill. If the query is highly
selective (retrieve, say, 1 GB or less of the 10 TB), then we can ignore must
of Drill's own overhead for now.
Assume you have good hard disks, with sustained 100 MB/sec. read performance.
It will thus take 10 TB / 100 MB/s = 10^13 / 10^8 = 10^5 = 10,000 seconds for a
single disk. Your budget is 5 * 60 = 300 seconds. So, you will need 10,000 /
300 = 34 disks working concurrently. Not a horrible first estimate.
Given that, you can then work out a hardware setup. If you are in the cloud,
you'll have to work out how to use cloud resources to get this throughput. If
on prem, then you probably have a standard hardware configuration. And, if you
have SSDs, the calculations will be different.
You'll want to test with plain HDFS (if that is what you're using) to figure
out the combination of disks-per-node to get the needed throughput. For
example, with 10 disks per node, you'd need the 4 nodes that Abhishek
recommended.
Maximum throughput depends on Drill being able to consume data at the desired
rate. Here you can do performance tests to see how fast a single Drill minor
fragment (thread) can read data on your hardware for your data. In my own tests
(on a Mac), I found something like 50 MB/s for a CSV file with a single
100-byte record. Your results will completely depend on record structure and
file format (CSV vs. JSON vs. Parquet.)
Once you know per-fragment throughput, you can target the needed number of
fragments to consume the disk I/O on each node. This will tell you the number
of cores. (Or, more realistically, given the number of available CPUs, how
much disk I/O can Drill consume on a single node?) If, say, you have 24 cores
and each can do 50 MB/sec, then your optimal Drill ingest rage is 1200 MB/sec,
which would, under ideal conditions, consume the output of 12 disks. Of course,
things are seldom optimum. The point is, balance disks and cores per node to
avoid bottlenecks.
Of course, you may not get the optimum sustained read rate on each disk for any
number of reasons (limited controller bandwidth, disk contention, HDFS
bottlenecks, etc.) So, having more nodes is better, with data spread across
more disks. If you get only 50 MB/s sustained, then you need, say, 70 disks. If
you have 16 cores per machine, and each can consume, say, 25 MB/s, then you can
have 8 disks per machine for a total of 70 / 8 = 9 machines.
The point is, make some assumptions then test them. Use the results to work out
an estimate of the number of disks and cores. Use testing to refine from there.
Thanks,
- Paul
On Tuesday, July 31, 2018, 2:12:17 AM PDT, Surneni Tilak
<[email protected]> wrote:
Hi Abhishek,
Thanks for your response . I will try with the approach that you have suggested
and come back if I I need any further help.
Best regards,
________
Tilak
-----Original Message-----
From: Abhishek Girish [mailto:[email protected]]
Sent: Monday, July 30, 2018 9:43 PM
To: user <[email protected]>
Subject: Re: Drill Configuration Requirements To Query Data in Tera Bytes
Hey Tilak,
We don't have any official sizing guidelines - for planning a Drill cluster. A
lot of it depends on the type of queries being executed (simple look-ups vs
complex joins), data format (columnar data such as Parquet shows best
performance), and system load (running a single query on nodes dedicated for
Drill).
It also depends on the type of machines you have - for example with beefy nodes
with lots of RAM and CPU, you'll need fewer number of nodes running Drill.
I would recommend getting started with a 4-10 node cluster with a good amount
of memory you can spare. And based on the results try and figure out your own
sizing guideline (either to add more nodes or increase memory [1]).
If you share more details, it could be possible to suggest more.
[1] http://drill.apache.org/docs/configuring-drill-memory/
On Mon, Jul 30, 2018 at 1:57 AM Surneni Tilak <[email protected]>
wrote:
> Hi Team,
>
> May I know the ideal configuration requirements to query data of size
> 10 TB with query time under 5 minutes. Please suggest me regarding the
> number of Drilbits that I have to use and the RAM(Direct-Memory &
> Heap_Memory) that each drill bit should consists of to complete the
> queries within the desired time.
>
> Best regards,
> _________
> Tilak
>
>
>
