Datasets:

---

openclimatefix

/

uk_pv

like 14

Follow

Open Climate Fix 104

The way the data is structured at the moment

The data in the parquet files is currently stored in a very "tall and thin" shape, like this:

Proposed data structure

I'd like to propose that we pivot the data, to get a very wide data structure like this:

Benefits of the pivoted structure

Benefit 1: Easier & faster to query

A large benefit is that the pivoted shape is easier to query (less code to write) and much faster to query (because the dataframe library doesn't have to scan through every row for every query). One of the nice features of Parquet files is that users can lazily open a large Parquet file and only load the columns that they need into memory. But users can only take advantage of this feature if the data is laid out such that uses can load only a subset of the columns. In the "tall and thin" shape, even if users only want to load data for a single PV system, users must still load all the data into memory.

Benefit 2: Smaller!

The pivoted shape requires a sixth of the storage space (which makes it easier to store and faster to download):

A year (2018) of 30-minutely PV data takes up this much space:

Compressed and using 64-bit numerical types:

Uncompressed (in memory) and using 64-bit numerical types:

Using more concise data types:

Why is the pivoted data structure smaller?

In the "tall and thin" shape, we repeat each SS_ID 17,520 times per year of data (the number of half hours in a year), and we repeat each datetime label 25,000 times. In constrast, in the pivoted structure, we're only storing each datetime label once, and we're only storing each SS_ID once.

Benefit 3: Parquet is columnar. So our data should be columnar.

In the "tall and thin" shape, we're not really storing data in the way Parquet is designed to store data. Parquet is a columnar format (see WikiPedia for a concise explanation of what "columnar" means). In short: Parquet works best when each column represents a contiguous, ordered sequence of measurements from a single object.

Disadvantages to using a pivoted structure

The only downside I can think of is that it's a different data structure, so people who are expecting the old data structure might get a shock! (And we'd have to change OCF's data prep code)

Conclusions

So, should we pivot all the Parquet files in this dataset?

Please let us know your thoughts!