Allow processing of duplicated attribute data columns

[olejandro]

Implications of duplicate labels for operations in pandas are described here: https://pandas.pydata.org/docs/user_guide/duplicates.html#consequences-of-duplicate-labels

This PR ensures the data in a duplicated column is returned only once (and not e.g. twice if 2 columns with the same name are present) during processing.

[SamRWest]

(edited - I misunderstood the logic here originally)
This doesn't throw an exception for austimes, but it seems to retains the duplicated columns, which then get turned into duplicate rows after the explode, e.g. the first row, which looks like this when passed into explode():

just gets exploded into this at the end of the function, retaining the duplicate fixom columns (I appended the name column for clarity):

This happens to be ok in this particular example because the values are identical, but if, say, the second fixom value was None, then dropping first duplicate rows would give the wrong answer later.

Whereas I understood that we want to want to 'fill-right' their values before exploding (e.g. using the merge function from #198) - so that any missing values in the right-most column get filled with non-missing values from the next column to the left, etc.

Suggest we replace this (assuming the attribute/alias stuff is sorted at this point) with:

if len(set(data_columns)) < len(data_columns):
    df =  _merge_duplicate_named_columns(df) #see #198

[olejandro]

@SamRWest, as I've tried to explain in #198, there may be an attribute alias, e.g.:

PROCESS	NCAP_START	START	NCAP_START
PRC1	2012		2020
PRC2	2012	2016
PRC3		2016	2020

Using your approach will turn this to:

PROCESS	START	NCAP_START
PRC1		2020
PRC2	2016	2012
PRC3	2016	2020

It will later be transformed into :

PROCESS	ATTRIBUTE	VALUE
PRC1	NCAP_START	2020
PRC2	NCAP_START	2016
PRC2	NCAP_START	2012
PRC3	NCAP_START	2016
PRC3	NCAP_START	2020

Resulting in an incorrect NCAP_START value for PRC2, because the rows below overwrite the rows above.

What I try to ensure with my changes is that the original table results in:

PROCESS	ATTRIBUTE	VALUE
PRC1	NCAP_START	2012
PRC1	NCAP_START	2020
PRC2	NCAP_START	2012
PRC2	NCAP_START	2016
PRC3	NCAP_START	2016
PRC3	NCAP_START	2020

This way the NCAP_START value for PRC2 will be correct.

[olejandro]

I think the approach proposed in #198 should be applied in normalize_column_aliases and only apply to columns not in config.all_attributes and config.attr_aliases sets.

[olejandro]

This happens to be ok in this particular example because the values are identical, but if, say, the second fixom value was None, then dropping first duplicate rows would give the wrong answer later.

@SamRWest, I've modified Demo 7 to cover this case. It seems to be working fine. Also we normally do drop rows with invalid values before doing any kind of overwritting.

[SamRWest]

Oh ok, I think I get it (didn't realise how aliases worked before). The explode() and later dropping of duplicate rows kind of does the work or resolving duplicates for you.

If you have time, it'd probably be worth writing a unit test to spell this logic out. But if you've checked that it matches VEDA, then that's the main thing, so I'm happy.

[olejandro]

Ok, sure! Tbh, I have never written one, but there is always the first time. 😃

[SamRWest]

Nothing to it, just make a fake dataframe (like your examples above) run it through the appropriate transforms, and assert what it should look like on the other end.
All the boilerplate is already set up in test_transforms.py, so just make a new test_whatever() function in there, and it'll get checked in every CI run so we don't accidentally change the desired behaviour later.

[olejandro]

I've udpated the benchmarks to cover a case like this:

PROCESS	NCAP_START	START	NCAP_START
PRC1	2012		2020
PRC2	2012	2016
PRC3		2016	2020

Currently the tests are failing on Demos7 due to additional records generated by an alias. The records appear in the correct order, so the input for GAMS will be correct.

[olejandro]

Now that the additional records are gone the tests seem to be failing due to failing updated Demo 7 and Demo 7r benchmarks when run on main. This makes sense because main cannot handle the type of duplicate columns introduced in Demo 7. @siddharth-krishna I guess we'll have to override the check?

[siddharth-krishna]

Failing tests: would it work if we didn't update the demos-xlsx ref in this PR, merged it in, and updated the benchmarks in the next one?

+1 for having a unit test with the nice example you made. (Perhaps that makes updating the benchmarks redundant? Maybe for future PRs it may be easier to update unit tests than the benchmarks repo)

[olejandro]

Thanks guys! Will revert the change to benchmarks version and add it together with a unit test in my next PR.