Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Allow optimistic mirror of kbuckets with read lock only #229

Merged
merged 4 commits into from
Feb 21, 2024

Conversation

emhane
Copy link
Contributor

@emhane emhane commented Jan 22, 2024

Description

allows optimistic and less blocking interaction with discv5

Notes & open questions

Change checklist

  • Self-review
  • Documentation updates if relevant
  • Tests if relevant

@emhane emhane force-pushed the rlock-on-kbuckets-in-api-call branch from e1cc1b9 to a29538d Compare January 22, 2024 20:10
Copy link
Member

@jxs jxs left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi Emilia, doesn't the kbuckets function cover these usecases?

@emhane
Copy link
Contributor Author

emhane commented Jan 25, 2024

Hi Emilia, doesn't the kbuckets function cover these usecases?

hey :) one of them, yes true, NodeStatus of entries would get cloned on cloning kbucktes. still, what else gets cloned? I want to return as little mem as possible from discv5 instance. so the node_ids only call is essential.

also it occurred to me, a try_r and try_w method in the api would be super useful. then we can implement async locks in the app using discv5. maybe you would like to implement this?

@jxs
Copy link
Member

jxs commented Jan 26, 2024

hey :) one of them, yes true, NodeStatus of entries would get cloned on cloning kbucktes. still, what else gets cloned? I want to return as little mem as possible from discv5 instance. so the node_ids only call is essential.

but table_entries_rlock also clones the entry, with the added disadvantage that you are iterating again just to collect after, to probably iterate again after right?

Regarding little as memory, a clone of KBucketsTable is returned with kbuckets which can last as little as a scope to be iterated and gathered the data required, without the double iteration. Can I ask you the reason for the memory constraints? To avoid the double iteration you can follow this technique and return impl Iterator<Item =NodeId>

@emhane
Copy link
Contributor Author

emhane commented Jan 28, 2024

hey :) one of them, yes true, NodeStatus of entries would get cloned on cloning kbucktes. still, what else gets cloned? I want to return as little mem as possible from discv5 instance. so the node_ids only call is essential.

but table_entries_rlock also clones the entry, with the added disadvantage that you are iterating again just to collect after, to probably iterate again after right?

but it only clones the node id (32 bytes) and not the enr (up to 300 bytes encoded length) for each entry

Regarding little as memory, a clone of KBucketsTable is returned with kbuckets which can last as little as a scope to be iterated and gathered the data required, without the double iteration. Can I ask you the reason for the memory constraints? To avoid the double iteration you can follow this technique and return impl Iterator<Item =NodeId>

I want to use this api in reth, using minimal resources is baseline for all client development.

/// .map(|entry| *entry.node.key.preimage())
/// .collect::<Vec<_>>());
/// ```
pub fn with_kbuckets<F, T>(&self, f: F) -> T
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

replaced the methods for this one which covers all cases

@emhane emhane requested a review from jxs January 28, 2024 17:34
@AgeManning
Copy link
Member

The idea is to not Clone the kbuckets table right?

I was thinking about this, and the data structure for kbuckets ins't great. But I estimate in most cases, we usually have a couple of buckets full, then the rest dying off. So we probably have like 50 ENRs in a table. If they are maxed out that's like 300 bytes per ENR which is like 15kb per table. So worst case we are prob copying 15kb in memory, which isn't too bad imo, but I can see the desire to avoid the clone.

I think in this case, an easy solution would be just to expose self.kbuckets as an Arc<RwLock> right? Because then you could do whatever you like to the Arc?

I think the original problem with exposing this, is that the public API is then tagged to the Arc and the exact version of parking_lot that exposes RwLock, so we'd have to re-export parking_lot. As this current PR exposes this now anyway, maybe we just expose self.kbuckets?

Will that satisfy all the requirements you need?

@emhane
Copy link
Contributor Author

emhane commented Jan 29, 2024

The idea is to not Clone the kbuckets table right?

I was thinking about this, and the data structure for kbuckets ins't great. But I estimate in most cases, we usually have a couple of buckets full, then the rest dying off. So we probably have like 50 ENRs in a table. If they are maxed out that's like 300 bytes per ENR which is like 15kb per table. So worst case we are prob copying 15kb in memory, which isn't too bad imo, but I can see the desire to avoid the clone.

I think in this case, an easy solution would be just to expose self.kbuckets as an Arc right? Because then you could do whatever you like to the Arc?

I think the original problem with exposing this, is that the public API is then tagged to the Arc and the exact version of parking_lot that exposes RwLock, so we'd have to re-export parking_lot. As this current PR exposes this now anyway, maybe we just expose self.kbuckets?

Will that satisfy all the requirements you need?

totally agree, I changed the code in this pr to do that yesterday

@AgeManning
Copy link
Member

Yeah, so what about something like this?:

    /// Returns the routing table of the discv5 service
    pub fn kbuckets_arc(&self) -> Arc<RwLock<KBucketsTable<NodeId, Enr>> {
        self.kbuckets.clone()
    }

@emhane
Copy link
Contributor Author

emhane commented Feb 1, 2024

Yeah, so what about something like this?:

    /// Returns the routing table of the discv5 service
    pub fn kbuckets_arc(&self) -> Arc<RwLock<KBucketsTable<NodeId, Enr>> {
        self.kbuckets.clone()
    }

the code I have made is essentially the same thing, besides it's safer, since any lock taken is at least dropped at the end of closure.

@AgeManning
Copy link
Member

Yeah, this api has a few downsides:

  • It exposes the specific version of parking_lot and Arc the library is using
  • It allows the user to potentially create dead-locks

I don't mind adding it in, if you really need it.

The downside of putting a closure into the API, is that its now limited to a FnOnce. If another user comes along that needs something more generic, Fn FnMut etc, then we'd have to create more APIs. Where-as just giving them the Arc, they can do with whatever they want with it. Its up to them to not cause deadlocks by holding locks imo.

I'm not sure other people should use this for the downsides above. But if you need it in reth, happy to add it in.

@emhane
Copy link
Contributor Author

emhane commented Feb 8, 2024

Yeah, this api has a few downsides:

  • It exposes the specific version of parking_lot and Arc the library is using

the same goes for the Enr type used by discv5, had some issues due to this. would have liked to set another const generic than the CombinedKey type, will make an issue to use this as default rather than the only type supported by discv5.

sure it exposes the specific parking lot version. this api allows to call try_read and try_write, I want that so I can make an async wrapper for it.

  • It allows the user to potentially create dead-locks

yes, In the docs I added a note on the user's responsibility to drop the lock. the risk is in practice minimised by locking in a closure's scope like this in my experience.

I don't mind adding it in, if you really need it.

it would be very useful to have that flexibility. I could even feature gate it if you like, have some feature called "dev-mode", wdyt?

The downside of putting a closure into the API, is that its now limited to a FnOnce. If another user comes along that needs something more generic, Fn FnMut etc, then we'd have to create more APIs. Where-as just giving them the Arc, they can do with whatever they want with it. Its up to them to not cause deadlocks by holding locks imo.

to take a read or write lock we just need a ref to self FnOnce. why would we need FnMut or Fn here?

I'm not sure other people should use this for the downsides above. But if you need it in reth, happy to add it in.

that would be great, thanks.

@AgeManning AgeManning merged commit 3c0d3d7 into sigp:master Feb 21, 2024
6 checks passed
@emhane emhane deleted the rlock-on-kbuckets-in-api-call branch March 2, 2024 23:43
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants