Principle:Danijar Dreamerv3 Distributed Logger Aggregation

Knowledge Sources	DreamerV3
Domains	Reinforcement_Learning, Distributed_Systems, Monitoring
Last Updated	2026-02-15 09:00 GMT

Overview

A centralized logging process that receives metrics and transitions from all distributed processes via RPC and writes aggregated statistics to configured output backends.

Description

In distributed DreamerV3 training, a dedicated logger process acts as a central aggregation point. It receives two types of RPC calls:

• add(metrics): Scalar metrics from the learner (training losses, FPS), replay (buffer stats, limiter stats), and environments (env FPS)
• tran(transitions): Batched transitions from the actor, used to compute episode statistics (scores, lengths, reward rates)

The logger maintains per-environment episode accumulators, handles episode timeouts for stale environments, and periodically writes all aggregated metrics to the configured backends (JSONL, TensorBoard, WandB, etc.).

Usage

The logger process runs independently in distributed mode. All other processes are its clients, sending metrics asynchronously. It is the single writer to all logging backends.

Theoretical Basis

Pseudo-code Logic:

# Abstract algorithm
server = RPCServer(logger_addr)
server.bind('add', lambda metrics: logger.add(metrics))
server.bind('tran', lambda transitions: process_episodes(transitions))

while True:
    if should_log():
        logger.write()  # Flush to all backends
    if should_save():
        checkpoint.save()  # Save step counter

Related Pages

Implemented By

• Implementation:Danijar_Dreamerv3_Parallel_Logger