tf.data.experimental.service.DispatcherConfig

View source on GitHub

Configuration class for tf.data service dispatchers.

tf.data.experimental.service.DispatcherConfig(
    port=0, protocol=None, work_dir=None, fault_tolerant_mode=False,
    job_gc_check_interval_ms=None, job_gc_timeout_ms=None
)

Fields:

• port: Specifies the port to bind to. A value of 0 indicates that the server may bind to any available port.
• protocol: The protocol to use for communicating with the tf.data service. Defaults to "grpc".
• work_dir: A directory to store dispatcher state in. This argument is required for the dispatcher to be able to recover from restarts.
• fault_tolerant_mode: Whether the dispatcher should write its state to a journal so that it can recover from restarts. Dispatcher state, including registered datasets and created jobs, is synchronously written to the journal before responding to RPCs. If True, work_dir must also be specified.
• job_gc_check_interval_ms: How often the dispatcher should scan through to delete old and unused jobs, in milliseconds. If not set, the runtime will select a reasonable default. A higher value will reduce load on the dispatcher, while a lower value will reduce the time it takes for the dispatcher to garbage collect expired jobs.
• job_gc_timeout_ms: How long a job needs to be unused before it becomes a candidate for garbage collection, in milliseconds. If not set, the runtime will select a reasonable default. A higher value will cause jobs to stay around longer with no consumers. This is useful if there is a large gap in time between when consumers read from the job. A lower value will reduce the time it takes to reclaim the resources from expired jobs.