v7.2.3.13

Changed

• Multi-head being turned off because cluster discovery failed is now
  reported as a warning; it was previously recorded only at debug level.
  Disabling auto-discovery deliberately is not a demotion and stays quiet.

Fixed

• GPUdbSqlIterator and DataFrameUtils.sql_to_df sharing one server-side
  paging table when built from a single options mapping. Used at the same
  time, the two raced to create it and one failed; where they did not overlap,
  either could return the other's rows instead of its own or fail with an
  index-out of-range error, and closing either dropped the table the other
  was still reading. Both now copy the options mapping and parameter list
  they are given. A result read in a single page was unaffected.
• One malformed address among those a worker rank advertises degrading the
  whole connection to head-node only, or failing it outright when a hostname
  regex was in use. The rank's other addresses are now tried, and the
  connection degrades only if none of them is usable.
• Multi-head ingestion sending records to a rank that had been removed from
  the cluster, which failed with an AttributeError. Ingestion and record
  retrieval now raise a GPUdbException naming the rank.
• Multi-head ingestion and retrieval sending records to the first worker rank
  instead of the head node, when a worker list was supplied and the head node
  was the one to be used--a replicated table, or auto-discovery turned off.
• A worker list built with use_head_node_only being routed as though it
  were a multi-head list, which failed a sharded insert or lookup with
  Sharded worker index is out of bound. The request is now honored.
• A record whose shard key held a negative zero or a NaN being sent to a
  different rank than the one the database assigns it.
• Sharded insertion and lookup sending every record to the first worker
  rank when no shard mapping was available. Records now go through the head
  node when the database cannot be reached, and the operation fails with an
  error naming the table when the database returns an empty mapping.

Notes

• Check CHANGELOG-FUNCTIONS.md for endpoint related changes