mirror of
https://github.com/Significant-Gravitas/AutoGPT.git
synced 2026-02-06 12:55:05 -05:00
d2a5bb286f
### Background
When multiple executors are executing the same node within the same graph execution, two node executions can read the same queue of input and read the same value—making the data that is supposed to be consumed once, consumed by two executions. The lack of lock & concurrency support for parallel execution within a single graph causes this issue.
Node concurrency also introduces poor UX in the current frontend implementation, when two nodes are executed in parallel, the current UI will not display its parallel execution update, but instead, it shows the updates that override each other. Until the execution observability is improved on the builder UI, this capability will be limited.
### Changes 🏗️
The scope of this change is to solve this issue by:
* Decouple Graph execution & Node execution, each has its own configured process pool.
* Make sure there is only 1 execution per node (we still allow parallel executions on different nodes) in a graph.
* Fixed concurrency issue by adding distributed lock API on agent_server.
* Few cleanups:
- Add more logging with geid & neid prefix on graph/node executions
- Moved execution status update to agent-server for a single source of status update (required by conn-manager/web-socket)
- Configured node parallelism to 10 & graph parallelism to 10 by default, so in the very rare worst-case, there can be 100 node executions.
- Re-use server resource for each integration test run