search

twosigma / flint

A Time Series Library for Apache Spark
Apache License 2.0
993 stars 184 forks source link

Unable to load Timeseries RDD from Parquet without sorting #59

Open d80tb7 opened 5 years ago

d80tb7 commented 5 years ago

Hi,

If I save a TimeseriesRDD to parquet as described in the documentation and then try to reread it without sorting then it will often fail with an error of the form: `"Partitions are not sorted. The partition 0 has the first key 1541810691000000000 and the partition 1 has the first key 1541780937000000000."

Doing a bit more research- I can see you have PR#16 which I think describes the issue in more detail and says that you are working around it using custom patched version of spark. I can see that the PR is year or so old, so presume it doesn't solve the issue?

If this is the case is there any other workaround that can be used so that you can save a timeseries rdd to disk and subsequently reload it without paying the cost of a sort?

`

dgrnbrg commented 4 years ago

Is there any update on this? I'd like to contribute ORC reader/writer functionality, but currently I need to include the sort operation when reading it into a flint DF, which isn't good enough to contribute I think.

icexelloss commented 4 years ago

Unfortunately #16 is left at an unfinished state. If someone can pick it up I am happy to review it but I am not sure when would I have time to do it myself.