[
https://issues.apache.org/jira/browse/PHOENIX-4229?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16194067#comment-16194067
]
Geoffrey Jacoby commented on PHOENIX-4229:
------------------------------------------
Thanks, [~tdsilva] - I'd looked at the link creation code earlier and missed
that the second link was created using CREATE_CHILD_LINK rather than a second
use of CREATE_VIEW_LINK. You're correct that there's only one Cell, LINK_TYPE,
that needs to be replicated, and it can be done purely on the basis of the
keys. I should be able to have a fix for this JIRA up in the next few days.
However, it looks like the child tenant id is in position 4 of the upsert
statement -- wouldn't that make it stored in System.Catalog.COLUMN_NAME?
COLUMN_FAMILY should be position 5, which holds the schema-qualified table
name. (It's not enough to know that this is a view by checking to see if
there's a parent table -- I need to know that it's a tenant-owned view.)
This all relies on an undocumented (even in code comments, as far as I can see)
assumption that System.Catalog.COLUMN_NAME becomes CHILD_TENANT_ID in this
circumstance, plus several other changed columns. Unless you know exactly what
to look for in the implementation itself, there's nothing to call out that any
column other than TENANT_ID has anything to do with tenant_ids.
As I said earlier, a document of System.Catalog's storage format, assumptions,
and use cases would be really helpful to have in these situations. It's not
just self-explanatory by reading the table's DDL, because System.Catalog is
logically half a dozen tables (several of them, like these rows, playing the
role of in-table indexes) all interleaved together.
> Parent-Child linking rows in System.Catalog break tenant view replication
> -------------------------------------------------------------------------
>
> Key: PHOENIX-4229
> URL: https://issues.apache.org/jira/browse/PHOENIX-4229
> Project: Phoenix
> Issue Type: Bug
> Affects Versions: 4.11.0, 4.12.0
> Reporter: Geoffrey Jacoby
> Assignee: Geoffrey Jacoby
>
> PHOENIX-2051 introduced new Parent-Child linking rows to System.Catalog that
> speed up view deletion. Unfortunately, this breaks assumptions in
> PHOENIX-3639, which gives a way to replicate tenant views from one cluster to
> another. (It assumes that all the metadata for a tenant view is owned by the
> tenant -- the linking rows are not.)
> PHOENIX-3639 was a workaround in the first place to the more fundamental
> design problem that Phoenix places the metadata for both table schemas --
> which should never be replicated -- in the same table and column family as
> the metadata for tenant views, which should be replicated.
> Note that the linking rows also make it more difficult to ever split these
> two datasets apart, as proposed in PHOENIX-3520.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
