Hidden service introduction point retries occur at 1 second intervals
Tor will try to reconnect to an introduction point up to 3 times. But rend_consider_services_intro_points() is called every second, which means that it uses up these retries very quickly, particularly if the connection fails quickly, like direct connections sometimes do on single onion services.
It might be more sensible to retry slightly more slowly.
On the other hand, maybe it's good that we fail fast and replace the introduction point.
This behaviour was introduced in commit 1125a4876b4.
Activity
changed milestone to %Tor: unspecified
Hrm... failing 3 times in 3 seconds means that every circuit creation failed right away? In that case, "tor" might have more problems... But I think the behavior here should be that we open a circuit and then if it fails like in 10 seconds after, we note down the try and retry a second later. That sounds reasonable to me?
Trac:
Status: new to needs_informationReplying to dgoulet:
Hrm... failing 3 times in 3 seconds means that every circuit creation failed right away? In that case, "tor" might have more problems... But I think the behavior here should be that we open a circuit and then if it fails like in 10 seconds after, we note down the try and retry a second later. That sounds reasonable to me?
Let's reword it to be something like:
When a circuit fails, we retry 10 seconds after we first detect the failure.
Then it's dynamic based on circuit failure time.
See #21621 (moved) for notes on the timing of the retries here: I think we should retry after 30 seconds to match the connection timeout (and avoid penalising slow hidden services).
To be clear, this ticket is about the onion service retrying circuits to its already-announced intro points, so it can resume using these intro points, so clients won't be too impacted when e.g. the onion service loses its network connection?
I ask because #25882 (moved) seems to be thinking this ticket is about clients who access onion services launching too many requests.
mentioned in issue #21621 (moved)
mentioned in issue #25882 (moved)
moved to tpo/core/tor#21600 (closed)
