What's the difference of FunctionDefHelper::Create and FunctionDefHelper::Define? - tensorflow

Question

the following is from math_grad.cc line543 to line554:
543 *g = FDH::Create("_",
544 // Input defs
545 {"x:T", "i:int32", "dy:T"},
546 // Ret val defs
547 {"dx:T", "di:int32"},
548 // Attr defs
549 {{"T: {half, float, double}"}},
550 // Nodes
551 nodes,
552 // Return values
553 {{"dx", "dx:output:0"}, {"di", "di:y:0"}});
554 return Status::OK();
and the following is from math_grad.cc line593 to line615,
593 Status MinMaxGradHelper(const string& op, const AttrSlice& attrs,
594 FunctionDef* g) {
595 // clang-format off
596 *g = FDH::Define(
597 // Arg defs
598 {"x:T", "i:int32", "dy:T"},
599 // Ret val defs
600 {"dx:T", "di:int32"},
601 // Attr defs
602 {{"T: {half, float, double}"}},
603 {
604 // keep_dims because we need to do x == y, which requires x
605 // and y are broadcastable.
606 {{"y"}, op, {"x", "i"}, {{"T", "$T"}, {"keep_dims", true}}},
607 {{"mask"}, "Equal", {"x", "y"}, {{"T", "$T"}}},
608 {{"mask_cast"}, "Cast", {"mask"}, {{"SrcT", DT_BOOL}, {"DstT", "$T"}}},
609 {{"mask_sum"}, "Sum", {"mask_cast", "i"}, {{"T", "$T"}}},
610 {{"norm_dy"}, "Div", {"dy", "mask_sum"}, {{"T", "$T"}}},
611 {{"sy"}, "Shape", {"y"}, {{"T", "$T"}}},
612 {{"norm_dy_reshaped"}, "Reshape", {"norm_dy", "sy"}, {{"T", "$T"}}},
613 {{"dx"}, "Mul", {"mask_cast", "norm_dy_reshaped"}, {{"T", "$T"}}},
614 {{"di"}, "ZerosLike", {"i"}, {{"T", DT_INT32}}}
615 });
I am a little confused about how does TensorFlow determine whether use
FDH::Define or FDH::Create.
As I know, FDH::Define is an old approch to define a FunctionDef. and I thought that FDH::Define provide less information compared with FDH::Create.
So, any one can help me tell the difference? Any advice will be appreciated, thanks a lot.

Answer 1

Answer from a colleague:
"FunctionDefs are not yet part of the public API; these are temporary mechanisms for defining functions until we have real API support for them.
One of those two functions is older and mirrors the older FunctionDef format, which has now been deleted.
I didn't delete the function, I just made it produce the new format; but this only works in simple cases. There was too much code using that function to migrate all of them.
But the main point is that this isn't documented well because it is not part of our public API and subject to change."