Rework "Normalize time series data intervals"

[amotl]

Hi there,

after reading @hammerhead's excellent Interpolation of missing time series values, I wonder if we should overhaul or even remove the Interpolate missing records section within the Normalize time series data intervals document.

Then, this document would be more like a Unix component in the terms of "do one thing and do it well" and I would rename it to "Processing and plotting timeseries-data with CrateDB, Pandas and Matplotlib" in order to better reflect its content.

What do you think about this?

With kind regards,
Andreas.

cc @proddata, @NOROSA

[proddata]

That was the intention, to maybe integrate it into the tutorial as next step :)

[nomicode]

I won't fight this work at all. the more tutorials the better

however, I don't think it's accurate to say that one is more about Pandas and Matplotlib and the other is (implicitly) more direct or otherwise somehow "closer" to the topic. both of them present the data in different ways. the new one uses screenshots of tables :) -- just another form of data representation

both are good! but I am not sure I would elevate one to over the other in terms of directness. the reason I went with graphical plotting is that I think it helps to see the data. I believe that most people are visual learners (but I may be misremembering the literature on this)

[proddata]

@NOROSA
What I discussed with @hammerhead was less about the data presentation, but rather of bringing in LOCF or NOCB into the normalize time series data tutorial, instead or additional to filling gaps with null 😉

[nomicode]

@proddata I see. that is a meaningful distinction yes! seems to me that these tutorials could be presented side by side as different-but-related tutorials

[hammerhead]

We can also remove the community article again at some later point in time if we decide to merge that information into the main documentation. For now, it was really just a quick way to get the information out to the public, because users ask about this topic frequently. I won't mind if anyone takes the content of the article and works it into the existing tutorial 👍.

I also just created a ticket to improve the backend support for time series interpolation. Depending on when/if this is going to be implemented, that might also be a good point in time to revisit the documentation situation.

[amotl]

Hi again,

do we have canonical tutorial-like resources, for example on the community forum or on the blog, which describe how to solve the LOCF/NOCB problem with CrateDB, using LAG/LEAD functions, which could be taken into consideration when modernizing the corresponding documentation sections?

I can see that a few query examples on the website are using LAG already, and that @andnig shared an example at 03-anomaly-detection.md:

Last Observation Carried Forward

More often than not, you are working with time series data using different sampling intervals. When
processing such data, you will most likely run into situations where you will have gaps in your data,
mostly represented by non-value symbols like NULL or NaN.

To fill these gaps, you can use CrateDB's LAG function, with their IGNORE NULLS option.

SELECT "time",
    COALESCE(battery_level, LAG(battery_level) IGNORE NULLS OVER w) AS battery_level,
    COALESCE(battery_status, LAG(battery_status) IGNORE NULLS OVER w) AS battery_status,
    COALESCE(battery_temperature, LAG(battery_temperature) IGNORE NULLS OVER w) AS battery_temperature
FROM machine_data
WINDOW w AS (ORDER BY "time");
ORDER BY "time";

Do you know about any other resources at cratedb.com, properly educating people about this?

With kind regards,
Andreas.

/cc @marijaselakovic, @karynzv, @hlcianfagna, @matriv, @seut