HN Explorer

Ask HN: Generating reports on time series data without killing performance?

5 points by Kendrick2 ↗ HN
I have an application which contains approximately 10 million time series records growing at a rate of 1 million per year. This application generates reports on subsets of the time series data. These reports have poor performance characteristics because for each record in a time series, another subset of time series data must often be loaded as part of the computations behind the report. Effectively this can lead to N^2 performance. The current workaround is to implement these reports as stored procedures in the Oracle database where the time series data resides. This saves network roundtrips for every time series record.

I've found the stored procedures to be insufficiently flexible to handle complex requirements compared to modern programming languages (requirements come in for new reports on a regular basis). I'd like to generate these reports in application code (C#) but can't see a way around the performance issue. Has anyone dealt with similar challenges and how did you work around them?

14 comments

[ 2.6 ms ] story [ 48.1 ms ] thread
eschutte2
That volume of data should be handled pretty easily by the right indexes and joins. At one place with that order of magnitude of data we kept an offline synced clone of the data with indexes designed for the queries we needed to run (this is basically OLAP). Have you already looked into that?
Kendrick2
Yes currently all the reports operate quite efficiently on a set of indexes. The performance of the database is currently very good with the reports in PL/SQL. The problem arises when the reports are generated using application logic on another server. Let's say we have a report for the last 10,000 records. For each of those records, I have to query the database again for some other time series information. If the latency between the reporting server and database server is 2ms, that report is going to take 20 seconds to generate.

Edit: If you're saying mirror the time series data onto the application server to avoid the network roundtrip, I guess that is possible but we still face the SQL overhead with each query.

eschutte2
Yeah, you definitely don't want to be making those roundtrips. Why doesn't a join work?
Kendrick2
Complex domain-specific requirements preclude simple joins unfortunately. This is why I'd like to redevelop these queries in C#. There are many imperative steps performed on each time series record to decide how these reports are generated.

To give a very broad example of one report: take 10,000 latest time series records from geographic location X. Now for each of those records load the latest 100 time series records from location Y and Z. Multiply all records from Y and Z together then divide by the value at location X. Throw in lots of business rules at every one of these steps so we have no chance of representing this as a tidy join operation and so that we can make no assumptions about the data we'll need to load to generate the report :)

flukus
Have you tried loading all the data required and then munging it in c#?

Edit - returning multiple result sets from the server is always an option too.

Kendrick2
> Have you tried loading all the data required...

Sorry, I probably edited my post while you were typing. The amount of data to load is theoretically unbounded (could load the entire data set for a big report) and it's very difficult to estimate ahead of time which data needs to be loaded. This is because the data in the time series dictates which secondary sets of data will be loaded on a record-by-record basis.

eschutte2
If it's not possible to do a few big queries to pull all the necessary base data (all the related Ys and Zs for the last 10,000 Xs) and then join/filter it yourself in code, then I think you are stuck with many queries. The options I can think of are: 1) keep the logic in stored procedures; 2) find a way to fetch the raw data in a few large queries, then join yourself; 3) let the reports be slow and generate them in the background; 4) write one complex SQL query that does all the crunching (if possible, you're saying it's not).
Kendrick2
> If it's not possible to do a few big queries...

Yeah, this is the unfortunate conclusion that I arrived at too. I'm looking at making these reports go through a job queue which e-mails the result to the user when it's done. This would allow me to develop much more complex, but much slower queries in C#.

maigrait
Are these java stored procedures?
tmaly
I am currently on a task of processing 4 million records to produce a report. It takes me about 10 seconds in the code.

You could always use a sliding window type algorithm to keep up with the grown of the data.

thorin
I think you're probably wrong that you can't handle it in stored procedures (packages). You may even be able to do a lot of the processing in a single SQL statement? Try this first, then PL/SQL. Do the bits you can't do in PL/SQL in a java class stored on the database...
brudgers
1. 10 million records is not big data.

2. If the rate of data acquisition is 1 million records a year, then 9 million records are the same as last year. There's no reason to hit the database again, because time series data [to be time series data] does not change.

3. This suggests that processing the data from the Oracle database is not a requirement...i.e. it could be moved to another system and processed there. Again, it's small data and doesn't change so duplication presents neither a storage issue nor a consistency issue.

4. The size and static nature of the data suggest that it might fit into memory on a single, moderately speced PC. An AWS type approach is also possible.

5. The right data storage format depends on the workload...maybe something column oriented?

Good luck.

ak39
If the requirements allow, consider replicating data in memory with SQLite :memory: option. 10 million records will be nothing for it.

Let us know.

NumberCruncher
When working with time series it is a good practice to move the data in memory and vectorise both the data and the data-manipulation logic. Doing so you can take advantage of the battle tested and fast linear algebra libraries like BLAS and LAPACK (or even move the calculation to the GPU). The last time I worked with time series data I used python-pandas. Even without former python knowledge it took only a couple of weeks to get productive. It has also a wrapper around SQLalchemy so you don't have to deal directly with the ORM.