Mark, That's a great question! 200 million per day equates to about 2K - 3K per second. So that is quite reasonable. You are very correct, though, that splitting that CSV into tons of one-line FlowFiles does indeed have a cost. Specifically, the big cost is the Provenance data that is generated at that rate. But again 2K - 3K per second going through a handful of Processors is a very reasonable workload.
I will caution you, though, that there is a ticket [1] where people will sometimes run into Out Of Memory Errors if they try to split a huge CSV into individual FlowFiles because it holes all of those FlowFile objects (not the data itself but the attributes) in memory until the session is committed. The workaround for this (until that ticket is completed) is to use a SplitText to split into 10,000 lines or so per FlowFile and then another SplitText to split each of those smaller ones into 1-line FlowFiles. Also of note, in 0.4.0, which is expected to be released in around a week or so, there is a new RouteText Processor. This, I think, will make your life far easier. Rather than using SplitText, Extract Text, and MergeContent in order to group the text, RouteText will allow you to supply a Grouping Regex. So that regex can just pull out the device id, year, month, and day, from each line and group together lines of text that have the same values into a single FlowFile. For instance, if your CSV looked like: # device_id, device_manufacturer, value, year, month, day, hour 1234, Famous Manufacturer, 83, 2015, 11, 13, 12 You could define a grouping regex as: (\d+), .*?, .*?, (\d+), (\d+), (\d+), .* It looks complex but it's just breaking apart the CSV into individual fields and grouping on device_id, year, month, day. This will also create a RouteText.Group attribute with the value "1234, 2015, 11, 13" This processor provides two benefits: it combines all of the grouping into a single Processor, and it cuts down on the millions of FlowFiles that are generated and then merged back together. As I write this, though, I am realizing that the regex above is quite a pain. We should have a RouteCSV processor as well. Though it won't provide any features that RouteText can't provide, it will make configuration far easier. I created a ticket for this here [2]. I'm not sure that it will make it into the 0.4.0 release, though. I hope this helps! Thanks -Mark [1] https://issues.apache.org/jira/browse/NIFI-1008 [2] https://issues.apache.org/jira/browse/NIFI-1161 <https://issues.apache.org/jira/browse/NIFI-1161> > On Nov 13, 2015, at 10:44 AM, Mark Petronic <[email protected]> wrote: > > I have a concept question. Say I have 10 GB of CSV files/day containing > records where 99% of them are from the last couple days but there are > stragglers that are late reported that can date back many months. Devices > that are powered off at times don't report but eventually do when powered on > and report old, captured data - store and forward kind of thing. I want to > capture all the data for historic reasons. There are about 200 million > records per day to process. I want to put this data in Hive tables that are > partitioned by year, month, and day and use ORC columnar storage. These > tables are external Hive tables and point to the directories where I want to > drop these files on HDFS, manually add new partitions, as needed, and > immediately be able to query using HQL. > > Nifi Concept: > > 0. Use GetText to get a CSV file > 1. Use UpdateAttribute to parse the incoming CSV file name to obtain a > device_id and set that as an attribute on the flow file > 2. Use SplitText to split each row into a flow file. > 3. Use ExtractText to identify the field that is the record timestamp and > create a year, month, and day attribute on the flow file from that timestamp. > 4. Use MergeContent with a grouping key made up of (device_id,year,month,day) > 5. Convert each file to ORC (many Parquet). This stage will likely require me > building a custom processor because the conversion is not going to be a > simple A-to-B. I want to do some validation on fields, type conversion, and > other custom stuff against some source-of-truth schema stored in a web > service with REST API. > 6. Use PutHDFS to store these ORC files in directories like > .../device_id/year=2015/month=11/day=9 by using the attributes already > present from the upstream processors to build up the path, where device_id is > the Hive table name and the year, month, day are the partition key name=value > per Hive format. The file names will just be some unique ID, they don't > really matter > 7. Use ExecuteProcessStream to execute a Hive script that will "alter table > add partitions...." for any partitions that were newly created on this > schedule run > > Is this insane or is it what Nifi was designed to do? I could definitely see > using a Spark job to do the group by (device_id,year,month,day) stage. 200M > flow files from the SplitText is the one that has me wondering if I am nuts > thinking of doing that? There must be overhead on flow files and deprecating > them to one line each seems to me as a worst case scenario. But it all > depends on the core design and whether Nifi is optimized to handle such a use > case. > > Thanks, > Mark
