* Not all "remote" repositories will fail. Specifically, if the remote
    repo uses the file:// protocol, and it doesn't refer to a shared
    filesystem, it will continue to be available.

  The question remaining is: Which level of offline mode will we support? It
  seems reasonable to assume that users will be able to tell when localhost is

You can check it at <a href="http://127.0.0.1:8000/docs" class="external-link" target="_blank">http://127.0.0.1:8000/docs</a>:

<img src="/img/tutorial/behind-a-proxy/image01.png">

But if we access the docs UI at the "official" URL using the proxy with port `9999`, at `/api/v1/docs`, it works correctly! 🎉

and it may store non-pointer data, C pointers, or Go pointers to pinned
memory through those pointers. It may not store a Go pointer to unpinned
memory in memory pointed to by a C pointer (which again, implies that it
may not store a string, slice, channel, and so forth). A Go function
called by C code may take a Go pointer but it must preserve the property
that the Go memory to which it points (and the Go memory to which that

            is ConeDefinitelyNotNullType -> original.getDirectSuperTypes(shouldApproximate).map {
                ConeDefinitelyNotNullType.create(it, analysisSession.useSiteSession.typeContext) ?: it
            }
            is ConeIntersectionType -> intersectedTypes.asSequence().flatMap { it.getDirectSuperTypes(shouldApproximate) }
            is ConeErrorType -> emptySequence()

!!! tip
    ...Although if the whole application is just **crashing immediately** it probably doesn't make sense to keep restarting it forever. But in those cases, you will probably notice it during development, or at least right after deployment.

    So let's focus on the main cases, where it could crash entirely in some particular cases **in the future**, and it still makes sense to restart it.

                        }
                    }
                    fir.getter?.let { getter ->
                        methods.singleOrNull { it.returnType != PsiType.VOID }?.let {
                            bindings.put(METHOD_FOR_FIR_FUNCTION, getter, Method(it.name, getAsmMethodSignatureWithCorrection(it)))
                        }
                    }

    for (auto it : op_it->second.input_tensor_id) {
      auto count_it = result.tensor_usage_counts.find(it);
      if (count_it != result.tensor_usage_counts.end()) {
        count_it->second++;
      } else {
        result.tensor_usage_counts[it] = 1;
        if (tensor_tape.find(it) != tensor_tape.end()) {
          tensor_stack.push_back(it);
        }
      }
    }

### Check the code

* Check and read the code, see if it makes sense, **run it locally** and see if it actually solves the problem.

* Then **comment** saying that you did that, that's how I will know you really checked it.

!!! info
    Unfortunately, I can't simply trust PRs that just have several approvals.

In the case of our dependency `get_settings()`, the function doesn't even take any arguments, so it always returns the same value.

That way, it behaves almost as if it was just a global variable. But as it uses a dependency function, then we can override it easily for testing.

        }

        val exitPoints = firTargets
            .mapNotNull { findLast(it) }
            .flatMap { node ->
                node.followingNodes
                    .filter { it !is StubNode }
                    .map { it.unwrap() }
                    .distinct()
                    .sortedBy { it.id }
            }.distinct()

        return exitPoints.size > 1
    }