As I said in the my initial message, precomputing is not an option.
Retrieving only the top/bottom N most correlated is an option – would that speed up the results?
Our SLAs are soft – slight variations (+- 15 seconds) will not cause issues.
Do you really need the results of all 3MM computations, or only the top- and bottom-most correlation coefficients? Could correlations be computed on a sample and from that estimate a distribution of coefficients? Would it make sense to precompute offline and instead focus on fast key-value retrieval, like ElasticSearch or ScyllaDB?
Spark is a compute framework rather than a serving backend, I don't think it's designed with retrieval SLAs in mind and you may find those SLAs difficult to maintain.
On Wed, Jul 17, 2019 at 3:14 PM Gautham Acharya <firstname.lastname@example.org> wrote:
Thanks for the reply, Bobby.
I’ve received notice that we can probably tolerate response times of up to 30 seconds. Would this be more manageable? 5 seconds was an initial ask, but 20-30 seconds is also a reasonable response time for our use case.
With the new SLA, do you think that we can easily perform this computation in spark?
Let's do a few quick rules of thumb to get an idea of what kind of processing power you will need in general to do what you want.
You need 3,000,000 ints by 50,000 rows. Each int is 4 bytes so that ends up being about 560 GB that you need to fully process in 5 seconds.
If you are reading this from spinning disks (which average about 80 MB/s) you would need at least 1,450 disks to just read the data in 5 seconds (that number can vary a lot depending on the storage format and your compression ratio).
If you are reading the data over a network (let's say 10GigE even though in practice you cannot get that in the cloud easily) you would need about 90 NICs just to read the data in 5 seconds, again depending on the compression ration this may be lower.
If you assume you have a cluster where it all fits in main memory and have cached all of the data in memory (which in practice I have seen on most modern systems at somewhere between 12 and 16 GB/sec) you would need between 7 and 10 machines just to read through the data once in 5 seconds. Spark also stores cached data compressed so you might need less as well.
All the numbers fit with things that spark should be able to handle, but a 5 second SLA is very tight for this amount of data.
Can you make this work with Spark? probably. Does spark have something built in that will make this fast and simple for you? I doubt it you have some very tight requirements and will likely have to write something custom to make it work the way you want.
On Thu, Jul 11, 2019 at 4:12 PM Steven Stetzler <email@example.com> wrote:
I am a beginner spark user too and I may not have a complete understanding of your question, but I thought I would start a discussion anyway. Have you looked into using Spark's built in Correlation function? (https://spark.apache.org/docs/latest/ml-statistics.html) This might let you get what you want (per-row correlation against the same matrix) without having to deal with parallelizing the computation yourself. Also, I think the question of how quick you can get your results is largely a data access question vs how fast is Spark question. As long as you can exploit data parallelism (i.e. you can partition up your data), Spark will give you a speedup. You can imagine that if you had a large machine with many cores and ~100 GB of RAM (e.g. a m5.12xlarge EC2 instance), you could fit your problem in main memory and perform your computation with thread based parallelism. This might get your result relatively quickly. For a dedicated application with well constrained memory and compute requirements, it might not be a bad option to do everything on one machine as well. Accessing an external database and distributing work over a large number of computers can add overhead that might be out of your control.
On Thu, Jul 11, 2019 at 9:24 AM Gautham Acharya <firstname.lastname@example.org> wrote:
Ping? I would really appreciate advice on this! Thank you!
This is my first email to this mailing list, so I apologize if I made any errors.
My team's going to be building an application and I'm investigating some options for distributed compute systems. We want to be performing computes on large matrices.
The requirements are as follows:
1. The matrices can be expected to be up to 50,000 columns x 3 million rows. The values are all integers (except for the row/column headers).
3. A sorted list of the correlation coefficients and their corresponding row keys need to be returned in under 5 seconds.
4. Users will eventually request random row/column subsets to run calculations on, so precomputing our coefficients is not an option. This needs to be done on request.
I've been looking at many compute solutions, but I'd consider Spark first due to the widespread use and community. I currently have my data loaded into Apache Hbase for a different scenario (random access of rows/columns). I’ve naively tired loading a dataframe from the CSV using a Spark instance hosted on AWS EMR, but getting the results for even a single correlation takes over 20 seconds.
Senior Data Scientist, Machine Learning Engineering
470 Park Ave South, 17th Floor, NYC 10016