`retry`: Don't retry on nxdomain

[wojtekmach]

[wojtekmach]

in fact the only TransportError we should retry on is timeout and econnrefused.

[wojtekmach]

We should also rename default mode :safe to :safe_transient. And add :transient which is :safe_transient except for all methods, including unsafe methods like PUT and POST too.

[wojtekmach]

In terms of matching on timeout and econnrefused, I think we should require exceptions returned from the steps to have a :reason field.

[wojtekmach]

Ref: https://github.com/wojtekmach/req/issues/144#issuecomment-1277753450

[dch]

this is actually desirable when connecting to a node where it is not known in advance whether it is ipv6 or not, via connect_options: [transport_opts: [inet6: true]]:

iex> Req.get("/_dbs")                                                                                                                             21:27:00.930 [error] retry: got exception, will retry in 1000ms, 3 attempts left

21:27:00.947 [error] ** (Mint.TransportError) non-existing domain
{:ok, %Req.Response{ status: 200, ...

We may get an nxdomain the first time, but succeed on the next pass, when we receive an ipv4 or ipv6 fallback from the previous attempt, via Finch / Mint.

[wojtekmach]

@dch oh I didn't realise we'd got nxdomain under those circumstances, definitely something to think about, thanks!

[whatyouhide]

You might also want to retry on econnrefused as another example, since it could be the server being stuffed or other temporary condition. In general I think it's a good idea to retry most transport errors, since most of them can be transient.