This page will serve to map out a general process for restoring/repairing content between nodes in Chronopolis, and denote areas where discussion/development is needed.
Chronopolis Repair Design Document
Michael Ritter, UMIACS October 10, 2016
Background
Within the standard operating of Chronopolis, it is likely due to the volume of data we ingest that
we will at some point need to repair data held at a node. In the event a node cannot repair their
own data, a process will be in place so that the data can be repaired through the Chronopolis
network. In this document a basic design proposal for a protocol through which we can repair
collections in a combination of manual and automated work will be outlined.
Considerations
As this design is still living, there are still open questions as to how everything should be finalized
and what impact they will have on the final result.
1. What transfer strategy should we use?
- Node to Node: Transfer between replicating nodes using rsync + ssh with no intermediary step
- Node to Ingest: Push content to the Ingest node from which a node can repair from
- ACE: Use ACE with https as the transfer mechanism for serving files
2. Should we create a new client for handling repair, or should it be merged in with the replication service?
- If it’s a new client, what type of application would be best (is cli good enough? do we want a gui? maybe some integration with the ingest server instead?)
3. Should we put a limit on the number of files being repaired in a single request?
4. Should we include tokens in this process, but leave implementation out for now?
Repair Flow
Basic flow: nodei = invalid; nodev = valid
1. nodei sees invalid files in ACEi
2. nodei gathers invalid files and issues a repair request to the ingest server
- This can be done in a standalone client
- Might want to consider having multiple requests in the event many files are corrupt
- POST /api/repair
3. nodev sees the repair request
4. nodev checks ACEv to see if the files are valid
- If the files are not valid, end this flow here
- Else: POST /api/repair/<id>/fulfill
5. nodev stages content for nodei
- if through ACE, create a token for nodei and make that available
- if p2p, make a link (or links) to the files in a home directory for nodei
- if through the ingest server, rsync the files up to the ingest server
6. nodev notifies content is ready for nodei
- POST /api/repair/fulfillment/<id>/ready
7. nodei replicates staged content
- GET /api/repair/fulfillment?to=nodei&status=ready
8. nodei issues an audit of the corrupt files
9. nodei responds with the result of the audit
- if the audit is not successful a new replication request will need to be made, but we might want to do that by hand
- POST /api/repair/fulfillment/<id>/complete
Transfer Strategies
Node to Node
Node to Node transfers would require additional setup on our servers, and would likely require
a look in to how we deal with security around our data access (transferring ssh keys, ensuring
access by nodes is read only, etc). A feasibly staging process could look like:
1. nodev links data (ln -s) in nodei’s home directory
2. nodei rsyncs data from nodev:/homes/nodei/depositor/repair-for-collection
Node to Ingest
Node to Ingest, while lengthy, would have the least amount of development and setup effort
associated with it. Since we will most likely not be repairing terabytes of data at a time, one
could say this is "good enough". The staging process for data would look similar to:
1. nodev rsyncs data to the ingest server
2. nodev notifies that the data is ready at /path/to/data on the ingest server
3. nodei rsyncs data from the ingest server on /path/to/data
ACE
Repairing through ACE would require additional development on ACE, as it currently does not
have any concept of API keys, but otherwise provides the same benefits of Node-to-Node repair
with some constraints from http itself. Staging would become quite simple, and amount to:
1. nodev marks the collection as allowing outside access (for API keys only?)
2. nodev requests a new temporary API key from ACE
3. nodei downloads from ACEv using the generated API key
API Design
The API can be viewed with additional formatting and examples at
http://adaptci01.umiacs.umd.edu:8080/
HTTP API
The REST API described follows standard conventions and is split in to two main parts, repair
and fulfillment.
Repair API
GET /api/repair/requests?<requested=?,collection=?,depositor=?,offers=?>
GET /api/repair/requests/<id>
POST /api/repair/requests
POST /api/repair/requests/<id>/fulfill
Fulfillment API
GET /api/repair/fulfillments?<to=?,from=?,status=?>
GET /api/repair/fulfillemnts/<id>
PUT /api/repair/fulfillments/<id>/ready
PUT /api/repair/fulfillemnts/<id>/complete
Models
A repair request, sent out by a node who notices they have corrupt files in a collection
Repair Request Model
{
"depositor": "depositor-with-corrupt-collection",
"collection": "collection-with-corrupt-files",
"files": ["file_0", "file_1", ..., "file_n"]
}
A repair structure, returned by the Ingest server after a repair request is received
Repair Model
{
"id": 1,
"status": "requested|fulfilling|repaired|failed",
"requester": "node-with-corrupt-file",
"depositor": "depositor-with-corrupt-collection",
"fulfillment": 3,
"collection": "collection-with-corrupt-files",
"files": ["file_0", "file_1", ..., "file_n"]
}
A fulfillment for a repair, returned by the Ingest server after a node notifies it can fulfill a repair
request. Credentials are only visible to the requesting node and administrators.
Fulfillment Model
{
"id": 3,
"to": "node-with-corrupt-file",
"from": "node-with-valid-file",
"status": "staging|ready|complete|failed",
"credentials": { ... }
"repair": 1
}
Credentials ACE
{
"type": "ace",
"api-key": "ace-api-key",
"url": "https://node_v/ace-am" # ?? Not sure if really needed
}
Credentials Node-to-Node
{
"type": "node-to-node",
"url": "node_i@node_v.edu:/homes/node_i/path/to/repair"
}
Credentials Node-to-Node
{
"type": "ingest",
"url": "node_i@chron.ucsd.edu:/path/to/repair"
}
-------------------------
Previous iterations:
Repair Design Document, October 2016
Previous repair discussion (labeled Restoration), June 2015
