Global
4. Global Configuration
Global configuration sets the execution environment. These settings can be overridden at the interface level or per processor.
storage: # where are files located?
...
compute: # what hardware / parallelism?
...
errors: # what happens on failures?
...
4.1 Storage
Configure the storage platform for datasets and outputs:
storage:
type: s3 # or local
bucket: datasets # required when type=s3
# Auth — choose one strategy
access_key: AKIAIOSFODNN7EXAMPLE
secret_key: wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY
session_token: <token> # temporary credentials
anonymous: false # anonymous access
role_arn: arn:aws:iam::123456789012:role/MyRole
session_name: dqm-ml # session name for role assumption
external_id: <external_id>
load_frequency: 900 # credential reload frequency (seconds)
# Connection
region: us-east-1
endpoint: https://s3.amazonaws.com # S3-compatible endpoint
request_timeout: 60 # seconds
connect_timeout: 5 # seconds
scheme: https
proxy_options: <url>
tls_ca_file_path: /path/to/ca.pem
# Retry
retry:
mode: standard # default | standard
max_attempts: 3
# Security
checksum_validation: when_required # when_required | always | never
4.2 Compute
Configure computation device, performance, and reproducibility:
compute:
seed: 42 # reproducibility seed
log_level: warning # debug | info | warning | error
device: cuda # or cpu
progress_bar: true # show tqdm progress bar during processing
threads: 4 # parallel reading threads
4.3 Errors (Global)
errors:
default: silent_fail # or fail_fast
max_failure_rate: 0.05 # auto-switch to fail_fast when exceeded
images:
on_decode_failure: silent_fail # corrupt bytes, missing file, S3 error
on_transform_error: silent_fail # resize / normalize / tensor conversion failure
on_unsupported_format: fail_fast # image format outside PIL's supported list
tabular:
on_missing_column: fail_fast
on_file_not_found: fail_fast
Interface-level and processor-level
errorsoverride global settings for their scope.