[
https://issues.apache.org/jira/browse/SOLR-10803?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16035800#comment-16035800
]
Hoss Man commented on SOLR-10803:
---------------------------------
Quick side digressions...
bq. Maybe also similar stuff to prevent FieldCache usage? ... I have seen that
Solr 7 allows to merge non-docvalues segments to ones with docvalues using
uninverter with a special mergepolicy. ...
That feels like a _very_ orthoginal idea (or 2?) that should really be
discussed in their own Jiras since they are broader in scope then just Trie vs
Point. (I'm not familar enough with what would be involved to even create the
jira)
bq. I'd suggest to also enable DocValues by default for all string/numeric/date
fields, unless explicitly disabled.
I like this idea -- but it's definitely orthoginal to the topic at hand.
I've spun that off into SOLR-10808
----
bq. I put the blocker priority since I think it is a better experience if all
7.x indices can be used with Solr 8, but there is also the possibility of just
removing Trie*Field in 8.0 and refusing to open any index that would make use
of those fields, even if they were created in 7.x.
Coincidently, sarowe & cassandra & I were just talking yesterday about our
concerns that beyond "known gaps" in terms of Solr features that work with Trie
fields but not (yet) Point fields (ex: SOLR-9989, SOLR-10503, SOLR-9985,
etc...) a larger concern as we move towards 7.0 is that test coverage of
PointFields in Solr is currently pretty shallow. We don't really have a very
good idea of what does/doesn't work with PointFields, which is disconcerting
for pushing them as the "default" (or "recommended") numeric types in Solr --
let alone forbiding the use of (new) Trie fields as suggested here.
Which is why I've started working on SOLR-10807 -- the current aim is a quick
and dirty way to help identify all of the potentially problematic areas as
quickly as possible, by forcing every test to use PointFields instead of
TrieFields. See comments in that Jira for details, but the nutshell is: at
this point it's hard to guess how many features/test might fail if we cut over
to PointFields -- because a big portion of our tests are using/expecting the
'id' field to be numeric, and before we can even get to the meat of the test,
using a Point based numeric as the 'id' field causes all sorts of problems
because they don't have any 'Terms' for updateDocument/deleteDocument
bq. In addition, the merge policy could also be used to convert Trie* to Point*
values by first uninverting (if no docvalues on trie) and redindexing the
fields during merging... (not sure how to do this, but should work somehow).
If we think it's viable to create a MergePolicy (Wrapper) that could convert
Trie fields to Point fields, then my straw many suggestion would be that in 7.x
we only discourage Trie fields, but not ban them completely -- with some
strongly worded warnings that Trie Fields will be completely removed in 8.0,
and any index that uses them will require manual upgrade using a special
converstion tool.
As things stand today, even if that MergePolicy/tool doesn't yet exist when 7.0
comes out, I'd rather say "7.x indexes using Trie fields *MAY* require
reindexing in 8.0, pending possible development of a tool to upgrade Trie
fields to Point fields" then ban Trie fields outright.
(Hell: As things stand today, even if we were confident it would be
_impossible_ to create such a tool/mergepolicy, I'd still rather say "7.x
indexes using Trie fields will *REQUIRE* reindexing in 8.0" then to change the
default configsets in 7.0 to use Points, let alone ban new Trie fields)
> Solr should refuse to create Trie*Field instances in 7.0 indices
> ----------------------------------------------------------------
>
> Key: SOLR-10803
> URL: https://issues.apache.org/jira/browse/SOLR-10803
> Project: Solr
> Issue Type: Bug
> Security Level: Public(Default Security Level. Issues are Public)
> Reporter: Adrien Grand
> Priority: Blocker
> Fix For: master (7.0)
>
>
> If we want to be able to remove support for legacy numerics from Solr in 8.0,
> we need to forbid the use of Trie*Field in indices that are created on or
> after 7.0.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
