refactor: Change concurrency approach #100

rklaehn · 2025-07-09T08:26:57Z

Description

Before, the concurrency approach for slots / handles for individual hashes was based on a top level task pool and "slots" that were managed in a map in the main actor. There were some tricky race conditions for the case where a handle would be "revived" when it was already executing its Drop fn - which does crucial things like writing the bitfield file. Also, there was actual parallelism per hash, which is probably not beneficial at all.

Now basically each hash runs effectively single threaded, meaning that we can later go from actual mutexes to more lightweight synchronisation primitives like https://crates.io/crates/atomic_refcell . Unfortunately everything must still be Send due to the fact that we run this whole thing on a multi-threaded executor 🤷 , thank you tokio. Otherwise we could just use a RefCell.

Now the concurrency is based on a task pool that will always contain at most a single task per hash. Multiple tasks that operate on the same hash are being handled concurrently, but not in parallel, using a FuturesUnordered. The drop case is handled in a cleaner way - when an actor becomes idle, it "gives back" its state to the owner - the manager actor.

If a task is being spawned while drop runs, these tasks go into the inbox and the actor gets revived immediately afterwards. The manager also has a pool of inactive actors to prevent reallocations.

All this is abstracted away by the entity_manager.

The entire entity_manager module could at some point become a separate crate, but for now I have inlined it so I don't need to do another crate.

Breaking Changes

Notes & open questions

Change checklist

Self-review.
Documentation updates following the style guide, if relevant.
Tests if relevant.
All breaking changes documented.

will probably inline this once it is proven to work well

futures-buffered is reexported by n0-future, as I learned yesterday

...by reverting a change that I did because of some older version of clippy

github-actions · 2025-07-09T08:32:34Z

Documentation for this PR has been generated and is available at: https://n0-computer.github.io/iroh-blobs/pr/100/docs/iroh_blobs/

Last updated: 2025-07-11T11:57:10Z

rklaehn · 2025-07-09T09:35:00Z

Here is where I am going with this. Now that we have the guarantee that each hash runs single threaded (not the same thread, thanks tokio, but at least 1 thread at a time), we can make concurrency much more lightweight: #101

rklaehn · 2025-07-09T10:04:00Z

src/store/fs.rs

            Command::ListBlobs(cmd) => {
                trace!("{cmd:?}");
-                let (tx, rx) = tokio::sync::oneshot::channel();


Unrelated change!

rklaehn · 2025-07-09T10:05:08Z

src/store/fs.rs

-        }
-        for hash in to_remove {
-            if let Some(slot) = self.handles.remove(&hash) {
-                // do a quick check if the handle has become alive in the meantime, and reinsert it


Getting rid of this nasty race condition handling was the whole purpose of the exercise. Also, we don't need a scheduled cleanup task anymore since the entity manager will deal with it. If an actor becomes idle it will be removed from the map after shutdown.

dignifiedquire · 2025-07-10T10:31:21Z

src/metrics.rs

@@ -4,6 +4,7 @@ use iroh_metrics::{Counter, MetricsGroup};

 /// Enum of metrics for the module
 #[allow(missing_docs)]
+#[allow(dead_code)]


why is this dead code?

I need to rethink what metrics to collect, many of those aren't even relevant anymore.

dignifiedquire · 2025-07-10T10:31:41Z

src/store/fs/util/entity_manager.rs

@@ -0,0 +1,1038 @@
+#![allow(dead_code)]


maybe remove this?

This thing was a separate crate. I inlined it so I don't have to publish and maintain yet another crate. The dead code part is the friendly API of the crate, if you don't want to embed the entity manager into your own actor.

But I want to embed the entity manager, since I don't want client -> iroh-blobs main actor -> entity manager -> entity. Save one hop.

src/store/fs/util/entity_manager.rs

dignifiedquire · 2025-07-10T10:37:23Z

src/store/fs/util/entity_manager.rs

+    }
+}
+
+/// A manager for entities identified by an entity id.


so you basically wrote a new actor system? 😅

No, that's irpc, in a way. (At least my idea how async boundaries should be structured). This is very specialized, so it's more like a task manager.

You get concurrency but not parallelism per entity, and parallelism between separate entities. An entity here is the state for a hash.

dignifiedquire · 2025-07-10T10:37:47Z

src/store/fs/gc.rs

@@ -299,6 +300,7 @@ mod tests {
            let outboard_path = options.outboard_path(&bh);
            let sizes_path = options.sizes_path(&bh);
            let bitfield_path = options.bitfield_path(&bh);
+            tokio::time::sleep(Duration::from_millis(100)).await; // allow for some time for the file to be written


this is very racy, isn't it?

Yeah. There is some delay between when an entity becomes idle and when the bitfield gets written to disk. The bitfield is ephemeral, but can be expensive to compute. You need to call shutdown before terminating the process to make sure all bitfields are persisted, but here we are looking for the file before shutting down, so there is currently no API to wait for it.

Not sure what to do about this. I could add an await_idle(hash). but that is not really a generic API, so it would have to be just for the FS store.

src/store/fs.rs

…y manager

rklaehn · 2025-07-10T15:32:09Z

One thing I am not sure about - the entity manager gives you an async but exclusive on_shutdown fn to perform cleanup like saving data to disk or an async io system.

Should it also do the same for waking up, have an async per entity cb that would be called on wakeup? I did not do it because I thought you might want to do this lazily or partially inside your state, so having 1 cb would be restrictive. But it sure would make the 2 test databases in the tests much more pleasant to write...

…ut a buggy system

src/store/fs/util/entity_manager.rs

rklaehn · 2025-07-11T12:06:48Z

One thing I am not sure about - the entity manager gives you an async but exclusive on_shutdown fn to perform cleanup like saving data to disk or an async io system.

Should it also do the same for waking up, have an async per entity cb that would be called on wakeup? I did not do it because I thought you might want to do this lazily or partially inside your state, so having 1 cb would be restrictive. But it sure would make the 2 test databases in the tests much more pleasant to write...

Nah, best handled by a tokio::sync::OnceCell or something custom.

rklaehn added 14 commits July 8, 2025 10:05

Add entity-manager dep

90c23be

will probably inline this once it is proven to work well

Remove unused dependencies

90db719

futures-buffered is reexported by n0-future, as I learned yesterday

Merge branch 'lose-futures-buffered-dep' into entity-manager

24c7f2c

Merge branch 'main' into entity-manager

126d834

Add import

264ed84

Add helpers for set and update

a859674

Make bipolar clippy happy again

baf4d54

...by reverting a change that I did because of some older version of clippy

Move message sending into meta

66f8540

Merge branch 'remove-db-public-send' into entity-manager

9193b08

WIP

8cae7cd

Replace slots with EntityHandler

161a70b

Get rid of the weak reference

169261a

Inline entity manager and streamline the dispatcher code a bit

6fe2e07

revert a few changes to minimize the diff

fd9faab

n0bot bot added this to iroh Jul 9, 2025

github-project-automation bot moved this to 🏗 In progress in iroh Jul 9, 2025

codespell

8ac1b78

remove debug code

011746d

rklaehn requested a review from Frando July 9, 2025 09:54

rklaehn commented Jul 9, 2025

View reviewed changes

more comments

def9abc

rklaehn force-pushed the entity-manager branch from 22a117b to def9abc Compare July 9, 2025 12:22

dignifiedquire reviewed Jul 10, 2025

View reviewed changes

src/store/fs/util/entity_manager.rs Show resolved Hide resolved

dignifiedquire reviewed Jul 10, 2025

View reviewed changes

src/store/fs.rs Show resolved Hide resolved

dignifiedquire reviewed Jul 10, 2025

View reviewed changes

src/store/fs.rs Show resolved Hide resolved

rklaehn added 5 commits July 10, 2025 17:34

Add property based tests for the entity manager

47cab9c

Add a property based test that tests basic functionality of the entit…

00c54f3

…y manager

clippy

664beb8

Merge branch 'main' into entity-manager

7b1b71b

fmt

e584ad1

rklaehn marked this pull request as ready for review July 10, 2025 15:29

rklaehn added 2 commits July 10, 2025 19:02

Add tests for the busy and dead cases which are hard to trigger witho…

4412be9

…ut a buggy system

clippy again

c26e555

matheus23 reviewed Jul 11, 2025

View reviewed changes

src/store/fs/util/entity_manager.rs Show resolved Hide resolved

Merge branch 'main' into entity-manager

8004115

refactor: Change concurrency approach #100

Are you sure you want to change the base?

refactor: Change concurrency approach #100

Uh oh!

Conversation

rklaehn commented Jul 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Breaking Changes

Notes & open questions

Change checklist

Uh oh!

github-actions bot commented Jul 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rklaehn commented Jul 9, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

rklaehn commented Jul 10, 2025

Uh oh!

Uh oh!

rklaehn commented Jul 11, 2025

Uh oh!

Uh oh!

rklaehn commented Jul 9, 2025 •

edited

Loading

github-actions bot commented Jul 9, 2025 •

edited

Loading