This talk is from October 2018, and so much has changed in the GOAI/RAPIDS ecosystem that it’s comical to see how much has changed! Regardless, the high-level concepts of how OmniSci works and the concepts behind GPU dataframes (then: pygdf, now: cudf) remain the same, so watching this talk still has value if you are interested in an end-to-end GPU workflow.
I’ve got PyData NYC 2018 in two days and rather finishing up my talk, I just realized that my source data has a silent corruption due to non-standard timestamps. Here’s how I fixed this using pandas and then uploaded the data to OmniSci.
Computers Are Dumb, MAKE THINGS EASIER FOR THEM!
Literally every data tool in the world can read the ISO-8601 timestamp format. Conversely, not every tool in the world can read Excel or whatever horrible other tool people use to generate the CSV files seen in the wild. While I should’ve been more diligent checking my data ingestion, I didn’t until I created a wonky report…
Let’s take a look at the format that tripped me up:
Month/Day/Year Hour:Minute:Second AM/PM feels very much like an Excel date format that you get when Excel is used as a display medium. Unfortunately, when you write CSV files like this, the next tool to read them has to understand 1) that these columns are timestamps and 2) if the user doesn’t specify the format, has to guess the format.
In my case, I didn’t do descriptive statistics on my timestamp columns and had a silent truncation(!) of the AM/PM portion of the data. So instead of having 24 hours in the day, the parser read the data as follows (the #AM and #PM are my comments for clarity):
So while the data looks like it was imported correctly (because, it is a timestamp), it wasn’t until I realized that hours 13-23 were missing from my data that I realized I had an error.
Pandas To The Rescue!
Fixing this issue is as straight-forward as reading the CSV into python using pandas and specifying the date format:
We can see from the code above that pandas has taken our directive about the format and it appears the data have been parsed correctly. A good secondary check here is that the difference in timestamps is -5, which is the offset of the East Coast of the United States relative to UTC.
Uploading to OmniSci Directly From Pandas
Since my PyData talk is going to be using OmniSci, I need to upload this corrected data or rebuild all my work (I’ll opt for fixing my source). Luckily, the pymapd package provides tight integration to an OmniSci database, providing a means of uploading the data directly from a pandas dataframe:
I have a pre-existing table hrl_load_metered on the database, so I can truncate the table to remove its (incorrect) data but keep the table structure. Then I can use load_table_columnar to insert the cleaned up data into my table and now my data is correct.
Computers May Be Dumb, But Humans Are Lazy
At the beginning, I joked that computers are dumb. Computers are just tools that do exactly what a human programs them to do, and really, it was my laziness that caused this data error. Luckily, I did catch this before my talk and the fix is pretty easy.
I’d like to say I’m going to remember to check my data going forward, but in reality, I’m just documenting this here for the next time I make the same, lazy mistake.
As my Developer Advocate role leads me to doing more and more Sysadmin/Data Engineer type of work, I continuously find myself looking for more efficient ways of copying data folders to where I need them. While there are a lot of great GUI ETL tools out there, for me the simplest and fastest way tends to be using linux utilities. Here’s how to mirror an FTP using lftp, with a cron repeater every five minutes.
Data are on an FTP, Need Further Processing
The problem I have is data that exists on a remote FTP, but are in a binary format that is incompatible with loading directly into OmniSci. My current plan is to use Python to convert the binary format into CSV, but with the data on a server that I don’t control, I need to make a copy somewhere else.
It’s also the case that the data are roughly 300GB per day, streaming in at various intervals across the day, so I need to make sure that any copying I do is thoughtful. Downloading 300GB of data per day is bad enough, doing it multiple times even worse!
lftp mirror to the Rescue!
The best choice in my case seems to be copying the files onto a VM I own. lftp has an option mirror to do just that. Here is the one-liner I’m using:
lftp -e : Execute command in quotes. In this case, the FTP allows anonymous access, so no user/pw arguments needed
mirror: Mirror command for lftp
-c: If download fails for whatever reason, keep trying (c = “continue”)
-e: Delete files on remote that are no longer on source (i.e. keep folders in perfect sync)
--parallel: Allow multiple connections for parallel downloading of multiple files
--verbose: Print lots of messages, helpful for debugging
With all of the flags in place, the last two arguments are the source (remote FTP) and destination (my VM) directories. Finally, I add a quit statement to exit lftp once the mirror process is over. This is mostly hygiene since I plan to run this on a cron scheduler and don’t want to leave the sessions open.
Run This Every Five Minutes, Forever
cron really is one of the greatest timesavers ever, especially in that it allows super-repetitive work to be automated away, usually with a single line. Here is the line I added after calling crontab -e on the command-line:
Quite simply, “every 5 minutes, run pull_from_ftp.sh”. Creating pull_from_ftp.sh is as straightforward as creating a text file:
That’s It? YES!
With just a few characters short of a full tweet, you can mirror an entire folder from an FTP, automatically. lftp in combination with cron helped me factor out hundreds of lines of pre-existing Python code, which not only removed untested, copy-pasted code from the workflow but also added parallel downloading, increasing data throughput.