Question: If you examine df after the plot has finished you will see that the content has changed since ternary changes it. If you feel that is not right and should be considered read-only memory then your ternary pyGMT wrapper needs to pass GMT_IS_DUPLICATE instead of GMT_IS_REFERENCE. You can make a decision on that but at least the SEGV bug in the API has been fixed, I hope.

Yes, I can confirm that this PR fixes the SEGV bug.

As what you said, the variable df changes from:

      a      b      c     d
0  0.16  0.331  0.509   9.0
1  0.22  0.625  0.158  39.0
2  0.09  0.180  0.732   2.0
3  0.66  0.258  0.078  12.0

to

          a         b     c     d
0  0.325500  0.286654   9.0   9.0
1  0.530907  0.539647  39.0  39.0
2  0.179641  0.155573   2.0   2.0
3  0.792169  0.224332  12.0  12.0

which is unexpected.

Yes, I do agree this is unexpected. Given that nobody will plot 2 Tb of points in a ternary (we hope) it is probably OK if ternary inside does the duplication to avoid surprises on the other end. But I think this is possibly also true for other situations. Many grid operators modify the input grid rather than duplicate, for instance.

Given that nobody will plot 2 Tb of points in a ternary (we hope) it is probably OK if ternary inside does the duplication to avoid surprises on the other end.

It would be better if ternary can do the duplication on the C side.