You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@cassandra.apache.org by "Sam Veloso (Jira)" <ji...@apache.org> on 2026/09/09 16:06:00 UTC

[jira] [Updated] (CASSCPP-17) Crash in update_indexes when adding an index for a table missing from schema metadata

     [ https://issues.apache.org/jira/browse/CASSCPP-17?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Sam Veloso updated CASSCPP-17:
------------------------------
    Description: 
During control-connection schema refresh (Cluster::on_reconnect → update_schema → update_indexes), the driver can dereference a null TableMetadata::Ptr and abort. This happens when system_schema.indexes contains two or more index rows for a table that is not present in the in-memory table metadata.
 
Observed with cpp-driver 2.17.1. The same unguarded add_index() is still present on master (src/metadata.cpp, Metadata::InternalData::update_indexes).
 
*Possible root cause*
update_indexes() already skips a missing table on the *first* index row, but it leaves table null while table_name is set to that missing table. Later index rows for the same table skip the lookup and call add_index() unconditionally:
 
{code:cpp}
if (table_name != temp_table_name) {
table_name = temp_table_name;
table = keyspace->get_table(table_name);
if (!table) continue;
table->clear_indexes();
}
 
table->add_index(IndexMetadata::from_row(index_name, buffer, row));
{code}
 
*Sequence:*
 # First index row for missing table T: get_table() fails, {{{}if (!table) continue{}}}, table stays null, table_name is T.
 # Next index row for the same T: {{table_name != temp_table_name}} is false, so the lookup is skipped.

{{table->add_index(...)}} runs on a null TableMetadata::Ptr.
 
A single leftover index row for a missing table does not crash (the continue already skips it). Two or more index rows for that table do.
 
This is inconsistent with the sibling helpers in the same file: update_columns() and update_legacy_indexes() wrap the mutation in {{{}if (table){}}}.
 
The window is realistic whenever tables and indexes are queried as separate result sets during reconnect, especially if a table was dropped (or not yet in metadata) while system_schema.indexes still has multiple indexes for it.
 
*Expected*
Index rows whose table is not in metadata should be skipped for every row, not only the first. The driver should not crash.
 
*Possible fix*
Guard add_index() the same way update_legacy_indexes() does:
 
{code:cpp}
if (table) {
table->add_index(IndexMetadata::from_row(index_name, buffer, row));
}
{code}
 
*Environment*
 * cpp-driver 2.17.1 (also current master)
 * Cassandra 3.x/4.x-style system_schema.indexes (update_indexes, not the legacy column-index path)
 * Trigger: control connection reconnect / schema metadata refresh while indexes exist for a table the driver does not have

  was:
During control-connection schema refresh (Cluster::on_reconnect → update_schema → update_indexes), the driver can dereference a null TableMetadata::Ptr and abort. This happens when system_schema.indexes contains two or more index rows for a table that is not present in the in-memory table metadata.
 
Observed with cpp-driver 2.17.1. The same unguarded add_index() is still present on master (src/metadata.cpp, Metadata::InternalData::update_indexes).
 
*Possible root cause*
update_indexes() already skips a missing table on the *first* index row, but it leaves table null while table_name is set to that missing table. Later index rows for the same table skip the lookup and call add_index() unconditionally:
 
{code:cpp}
if (table_name != temp_table_name) {
table_name = temp_table_name;
table = keyspace->get_table(table_name);
if (!table) continue;
table->clear_indexes();
}
 
table->add_index(IndexMetadata::from_row(index_name, buffer, row));
{code}
 
*Sequence:*
# First index row for missing table T: get_table() fails, \{{if (!table) continue}}, table stays null, table_name is T.
# Next index row for the same T: \{{table_name != temp_table_name}} is false, so the lookup is skipped.
# \{{table->add_index(...)}} runs on a null TableMetadata::Ptr.
 
A single leftover index row for a missing table does not crash (the continue already skips it). Two or more index rows for that table do.
 
This is inconsistent with the sibling helpers in the same file: update_columns() and update_legacy_indexes() wrap the mutation in \{{if (table)}}.
 
The window is realistic whenever tables and indexes are queried as separate result sets during reconnect, especially if a table was dropped (or not yet in metadata) while system_schema.indexes still has multiple indexes for it.
 
*Expected*
Index rows whose table is not in metadata should be skipped for every row, not only the first. The driver should not crash.
 
*Possible fix*
Guard add_index() the same way update_legacy_indexes() does:
 
{code:cpp}
if (table) {
table->add_index(IndexMetadata::from_row(index_name, buffer, row));
}
{code}
 
*Environment*
* cpp-driver 2.17.1 (also current master)
* Cassandra 3.x/4.x-style system_schema.indexes (update_indexes, not the legacy column-index path)
* Trigger: control connection reconnect / schema metadata refresh while indexes exist for a table the driver does not have


> Crash in update_indexes when adding an index for a table missing from schema metadata
> -------------------------------------------------------------------------------------
>
>                 Key: CASSCPP-17
>                 URL: https://issues.apache.org/jira/browse/CASSCPP-17
>             Project: Apache Cassandra C/C++ driver
>          Issue Type: Bug
>            Reporter: Sam Veloso
>            Priority: Normal
>
> During control-connection schema refresh (Cluster::on_reconnect → update_schema → update_indexes), the driver can dereference a null TableMetadata::Ptr and abort. This happens when system_schema.indexes contains two or more index rows for a table that is not present in the in-memory table metadata.
>  
> Observed with cpp-driver 2.17.1. The same unguarded add_index() is still present on master (src/metadata.cpp, Metadata::InternalData::update_indexes).
>  
> *Possible root cause*
> update_indexes() already skips a missing table on the *first* index row, but it leaves table null while table_name is set to that missing table. Later index rows for the same table skip the lookup and call add_index() unconditionally:
>  
> {code:cpp}
> if (table_name != temp_table_name) {
> table_name = temp_table_name;
> table = keyspace->get_table(table_name);
> if (!table) continue;
> table->clear_indexes();
> }
>  
> table->add_index(IndexMetadata::from_row(index_name, buffer, row));
> {code}
>  
> *Sequence:*
>  # First index row for missing table T: get_table() fails, {{{}if (!table) continue{}}}, table stays null, table_name is T.
>  # Next index row for the same T: {{table_name != temp_table_name}} is false, so the lookup is skipped.
> {{table->add_index(...)}} runs on a null TableMetadata::Ptr.
>  
> A single leftover index row for a missing table does not crash (the continue already skips it). Two or more index rows for that table do.
>  
> This is inconsistent with the sibling helpers in the same file: update_columns() and update_legacy_indexes() wrap the mutation in {{{}if (table){}}}.
>  
> The window is realistic whenever tables and indexes are queried as separate result sets during reconnect, especially if a table was dropped (or not yet in metadata) while system_schema.indexes still has multiple indexes for it.
>  
> *Expected*
> Index rows whose table is not in metadata should be skipped for every row, not only the first. The driver should not crash.
>  
> *Possible fix*
> Guard add_index() the same way update_legacy_indexes() does:
>  
> {code:cpp}
> if (table) {
> table->add_index(IndexMetadata::from_row(index_name, buffer, row));
> }
> {code}
>  
> *Environment*
>  * cpp-driver 2.17.1 (also current master)
>  * Cassandra 3.x/4.x-style system_schema.indexes (update_indexes, not the legacy column-index path)
>  * Trigger: control connection reconnect / schema metadata refresh while indexes exist for a table the driver does not have



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@cassandra.apache.org
For additional commands, e-mail: commits-help@cassandra.apache.org