[
https://issues.apache.org/jira/browse/NIFI-2854?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15663982#comment-15663982
]
ASF GitHub Bot commented on NIFI-2854:
--------------------------------------
Github user markap14 commented on the issue:
https://github.com/apache/nifi/pull/1202
@joshelser I did actually look into using both Protocol Buffers as well as
Avro to perform the serialization/deserialization. That really would be
preferred, as they are both very stable libraries and much more
"robust"/feature-rich than what we have here. Unfortunately, though, because of
the way that their readers/writers work, using those would have required some
pretty intense refactoring of some of the core repository code. This is largely
due to the API that was created for the repository wasn't thought through well
enough. For example, the RecordWriter has a `writeRecord` method that takes in
a record write as well as the OutputStream to write to. The repository itself
may write to the OutputStream in between records. These libraries wouldn't
really support that well. So rather than rewrite some of the most critical
parts of NiFi, I elected to create a much simpler schema-based reader/writer
approach. It may make sense at some point to review this decision though and
refactor the repositories to make them more amenable to this type of thing.
> Enable repositories to support upgrades and rollback in well defined scenarios
> ------------------------------------------------------------------------------
>
> Key: NIFI-2854
> URL: https://issues.apache.org/jira/browse/NIFI-2854
> Project: Apache NiFi
> Issue Type: Improvement
> Components: Core Framework
> Reporter: Mark Payne
> Assignee: Mark Payne
> Fix For: 1.1.0
>
>
> The flowfile, swapfile, provenance, and content repositories play a very
> important roll in NiFi's ability to be safely upgraded and rolled back. We
> need to have well documented behaviors, designs, and version adherence so
> that users can safely rely on these mechanisms.
> Once this is formalized and in place we should update our versioning guidance
> to reflect this as well.
> The following would be true from NiFi 1.2.0 onward
> * No changes to how the repositories are persisted to disk can be made which
> will break forward/backward compatibility and specifically this means that
> things like the way each is serialized to disk cannot change.
> * If changes are made which impact forward or backward compatibility they
> should be reserved for major releases only and should include a utility to
> help users with pre-existing data convert from some older format to the newer
> format. It may not be feasible to have rollback on major releases.
> * The content repository should not be changed within a major release cycle
> in any way that will harm forward or backward compatibility.
> * The flow file repository can change in that new fields can be added to
> existing write ahead log record types but no fields can be removed nor can
> any new types be added. Once a field is considered required it must remain
> required. Changes may only be made across minor version changes - not
> incremental.
> * Swap File storage should follow very similar rules to the flow file
> repository. Adding a schema to the swap file header may allow some variation
> there but the variation should only be hints to optimize how they're
> processed and not change their behavior otherwise. Changes are only permitted
> during minor version releases.
> * Provenance repository changes are only permitted during minor version
> releases. These changes may include adding or removing fields from existing
> event types. If a field is considered required it must always be considered
> required. If a field is removed then it must not be a required field and
> there must be a sensible default an older version could use if that value is
> not found in new data once rolled back. New event types may be added.
> Fields or event types not known to older version, if seen after a rollback,
> will simply be ignored.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
