User access logs
DACP nodes write completed Flight requests as JSON Lines for line-oriented collection. Logs are independent of execution history and do not retain stream contents. The default directory is .dacp-access/<port>/ relative to the server working directory, with a 64 MiB file limit.
access_log:
directory: /var/log/dacp/access
max_bytes: 67108864
max_bytes must be a positive integer. If the next record would exceed the limit of a nonempty file, a new file is opened before appending; one JSON record never spans files. An oversized record occupies its own file. Names are access-<timestamp>-<UUID>.jsonl; restarts never overwrite old logs. Old files are retained without automatic deletion; deployment tooling owns collection and retention. Request threads enqueue records in a bounded queue of at most 1024 entries, then return. A dedicated writer appends them serially without making each request wait for disk synchronization. The writer synchronizes every second or after 256 records, and before rotation and orderly shutdown. A full queue applies backpressure rather than silently dropping records. Disk write failures are logged server-side and surfaced on subsequent appends and shutdown. Abnormal termination may lose records not yet written or synchronized; access logs are not per-request transaction commits.
Fields include request ID, start/end timestamps, TLS-verified identity fingerprint, operation, logical Dataset/Item/view where applicable, status, elapsed time, sent batch rows and bytes. Bytes views count raw binary cell payloads; records bytes sum Arrow vector buffers and are not network wire traffic. Step exchanges count outer frames. Requests without a started stream have zero traffic.
Statuses include completed, failed, denied, and cancelled. Discovery, metadata, data streams, step exchanges and Union actions are logged. TLS handshake failures happen before Flight request dispatch and require transport/proxy logging. Entries omit raw tickets, operator source, parameters, data content, credentials, physical backend addresses, and raw exception messages.
The configured directory must be writable. Startup initializes logging; orderly shutdown drains the queue, synchronizes it, and releases file handles. Use execution history for step states and events; request completion does not establish business-job success.
Business operation names
operation uses business names rather than transport methods. Records are written when requests finish; a start timestamp does not mean a record was emitted on arrival. The separate transport security audit, authentication-audit.jsonl, still records RPC methods.
| Operation | Meaning |
|---|---|
catalog.list |
List datasets |
catalog.describe |
Catalog snapshot |
item.list |
List items |
item.describe |
Item metadata |
item.count |
Record count |
item.read |
Read an Item view |
node.list |
List available nodes |
job.events.read |
Read Job events |
job.result.read |
Read Job results |
job.submit |
Submit and execute |
step.infer |
Infer step outputs |
job.infer |
Infer Job outputs |
service.capabilities |
Service capabilities |
union.membership.get |
Union membership |
node.status |
Node resources and capacity |
job.prepare |
Prepare sub-job |
job.start |
Start sub-job |
job.status |
Execution status |
job.cancel |
Cancel execution |
job.commit |
Job result commit |
job.checkpoints.release |
Release execution-scoped checkpoints |
step.output.read |
Read cross-node step output |
request.unsupported |
Unsupported control request |
Exception diagnostics
Server runtime logs include full exception stacks and cause chains, with business operation, request_id, and the known job_id. Before a Job is created, its ID is marked unavailable. Background execution uses correlation identifiers; dispatched sub-job execution IDs correlate with Union execution history. Access logs retain outcomes and correlation identifiers without exception stacks, request bodies, or credentials. Job execution errors return a concise description and an available request ID to clients. Runtime logs are for operators: exception stacks can contain internal file paths and third-party error details.