Extensional equality is an equality such that when forall x, f x = g x then f = g

ConorMcBride noted that the following four properties cannot be satisfied simultaneously

1. Extensional equality
2. Intensional inductive families
3. Canonicity
4. Decidable type checking

Consider the following inductive family.

Inductive NatId : (nat -> nat) -> Prop :=
 mkNatId : NatId (fun x:nat => x).

Suppose that Coq had Extensional equality. This would mean that there would be theorems such as Zero_Right_Id : (fun x:nat => x+0) = (fun x:nat => x). This could be used to make an instance of the NatId family.

Definition mkNatId2 : NatId (fun x:nat => x+0) :=
 (eq_ind_r NatId mkNatId Zero_Right_Id).

Because mkNatId2 is a closed term of an inductive type, by canonicity its normal form must begin with a constructor. However, NatId only has one constructor therefore mkNatId2 must reduce to mkNatId and therefore their types must also be convertible

NatId (fun x:nat => x+0) <===> NatId (fun x:nat => x)

and therefore (fun x:nat => x+0) and (fun x:nat => x) must be convertible. In general, this convertibility check must work for any two extensionally equal functions and that is undecidable in general.

Coq chooses to not have extensional equality by default. Users can add an exentensionality axiom to get extensional equality by losing canonicity. Other proof systems may make other choices. For instance, Epigram 2 has no intensional inductive families, while other system may drop decidable type checking.

1. Epigram 2 will have inductive families upto extensional equality which is the natural choice for a system with extensional equality.
2. In practice all common inductive families I know (e.g. Vectors, finite types, typed lambda terms) are indexed over first order types and for them there is no difference between intensional or extensional families. The only important exception is equality also in the disguise of a singleton family.
3. Thierry Coquand already observed that pattern matching over intensional families is inconsistent with extensionality in the mid 90ies.