This makes sense to me.
I think the Create case was more obvious since it usually impacts time to readiness. For resources that do not require a provisioning time, e.g. a bucket, we want to mark it as Ready as soon as possible after the Create call, so, we re-queued immediately instead of waiting for the next reconcile (usually after the poll interval, assuming we don't update the CR in Create).

Reflecting on the Updates feels less of an issue compared to Create, but it is still valid.
We need to make sure it doesn't introduce loops (for whatever reason I can think of) or extra loads, at least by testing with some of the providers/resources out there.

Would this also mean that a failed update results in the resource no longer being ready? I think that would be really valuable; the fact that this doesn't happen (at least not always) today makes troubleshooting much harder.

Yes, it would work like that.

When the external.Update function returns an error, we can just add a xpv1.Updating() condition to the SetConditions call. This would set the resource as Ready: false.

update, err := external.Update(externalCtx, managed)
if err != nil {
    log.Debug("Cannot update external resource")
    record.Event(managed, event.Warning(reasonCannotUpdate, err))
    managed.SetConditions(xpv1.Updating(), xpv1.ReconcileError(errors.Wrap(err, errReconcileUpdate)))
    return reconcile.Result{Requeue: true}, errors.Wrap(r.client.Status().Update(ctx, managed), errUpdateManagedStatus)
}

I was thinking of opening a PR for this, but haven't had the time yet. It would require some testing too, but I am not that up to date on the code here.

Does a resource being updated mean it's not ready? I don't think so, at least not usually.

The Ready status condition indicates that the resource in the external system that the MR represents is ready to use.

When the provider requests that a resource be created, it typically takes a little time for the creation process to complete. During this time the resource isn't usable. The provider confirms the creation has succeeded before marking the resource ready.

When the provider requests that a resource be updated, it doesn't necessarily prevent anyone from using the external resource. Some resources (for example a GKE cluster) may enter a state like 'Repairing' and be degraded or unavailable while the external system processes the update. This isn't always the case though - in fact anecdotally I think it's less common.

Does a resource being updated mean it's not ready? I don't think so, at least not usually.

The Ready status condition indicates that the resource in the external system that the MR represents is ready to use.

When the provider requests that a resource be created, it typically takes a little time for the creation process to complete. During this time the resource isn't usable. The provider confirms the creation has succeeded before marking the resource ready.

When the provider requests that a resource be updated, it doesn't necessarily prevent anyone from using the external resource. Some resources (for example a GKE cluster) may enter a state like 'Repairing' and be degraded or unavailable while the external system processes the update. This isn't always the case though - in fact anecdotally I think it's less common.

I totally agree with this, we currently use the ready status on a resource to determine whether or not we can access status properties on the resource, which usually includes some form of ID / ARN.

From our perspective, a resource that is updating is still "ready" even if the update fails because its existing status properties remain valid.

For example we may request to expand a disk on a VM which fails, but other resources consuming that VM's ID should consider the VM ready.

Synced being set to false, might make sense, or an addition of a new state property.

More robust status conditions like it sounds like are being discussed here would be a better solution to crossplane/upjet#327 than what I proposed in that issue.