phoenix-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "James Taylor (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (PHOENIX-3823) Force cache update on MetaDataEntityNotFoundException
Date Wed, 17 May 2017 22:12:04 GMT

    [ https://issues.apache.org/jira/browse/PHOENIX-3823?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16014845#comment-16014845
] 

James Taylor commented on PHOENIX-3823:
---------------------------------------

Thanks for the patch, [~sukunaidu@gmail.com]. It's looking good. Here's some feedback:
* In UpsertCompiler, let's remove the while/catch/retry loop as you have it at a higher level
now instead:
{code}
            } catch (MetaDataEntityNotFoundException e) {
                // Catch column/column family not found exception, as our meta data may
                // be out of sync. Update the cache once and retry if we were out of sync.
                // Otherwise throw, as we'll just get the same error next time.
                if (retryOnce) {
                    retryOnce = false;
                    if (new MetaDataClient(connection).updateCache(schemaName, tableName).wasUpdated())
{
                        continue;
                    }
                }
                throw e;
            }
            break;
{code}
* Instead of checking for equality of the upsert and select TableNode (which won't be reliable
when dynamic columns are used and won't detect the case of a joined table matching the upsert
table) as you're doing below, we should instead pass in the table being upserted into (i.e.
{{table}}) as an extra argument to the FromResolver.getResolverForQuery(). Then in FromCompiler.createTableRef(),
always hit the server if the {{table}} schema and table name equal the table and schema name
being resolved.
{code}
+                    boolean alwaysHitServer = false;
+                    //Ignore cache when we do Upsert Select on same table
+                    if(tableNode!=null && tableNode.equals(select.getFrom())){
+                        alwaysHitServer = true;
+                    }
+                    ColumnResolver selectResolver = FromCompiler.getResolverForQuery(select,
connection, alwaysHitServer);
{code}
* The same issue exists in DeleteCompiler (though it's a bit more of an edge case - a sub
query in a DELETE referring to the same table as being deleted from). Based on the above,
you'd just need to pass in {{table}} to the FromResolver.getResolverForQuery() here:
{code}
                if (transformedSelect != select) {
                    resolverToBe = FromCompiler.getResolverForQuery(transformedSelect, connection);
                    select = StatementNormalizer.normalize(transformedSelect, resolverToBe);
                }
{code}


> Force cache update on MetaDataEntityNotFoundException 
> ------------------------------------------------------
>
>                 Key: PHOENIX-3823
>                 URL: https://issues.apache.org/jira/browse/PHOENIX-3823
>             Project: Phoenix
>          Issue Type: Sub-task
>    Affects Versions: 4.10.0
>            Reporter: James Taylor
>            Assignee: Maddineni Sukumar
>             Fix For: 4.11.0
>
>         Attachments: PHOENIX-3823.patch, PHOENIX-3823.v2.patch, PHOENIX-3823.v3.patch,
PHOENIX-3823.v4.patch, PHOENIX-3823.v5.patch
>
>
> When UPDATE_CACHE_FREQUENCY is used, clients will cache metadata for a period of time
which may cause the schema being used to become stale. If another client adds a column or
a new table or view, other clients won't see it. As a result, the client will get a MetaDataEntityNotFoundException.
Instead of bubbling this up, we should retry after forcing a cache update on the tables involved
in the query.
> The above works well for references to entities that don't yet exist. However, we cannot
detect when some entities are referred to which no longer exists until the cache expires.
An exception is if a physical table is dropped which would be detected immediately, however
we would allow queries and updates to columns which have been dropped until the cache entry
expires (which seems like a reasonable tradeoff IMHO. In addition, we won't start using indexes
on tables until the cache expires.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

Mime
View raw message