| The Greenplum pg_upgrade code has been extended to work with |
| Apache Cloudberry. This README is intended to document the changes |
| that have been done. |
| |
| Operation of pg_upgrade remains unchanged as are the available command |
| line parameters. Upgrading a Apache Cloudberry cluster with |
| pg_upgrade is done by first upgrading the QD and then the QEs. |
| |
| |
| Distributed Apache Cloudberry upgrades |
| ------------------------------ |
| |
| Apache Cloudberry requires that all Oids are synchronized between |
| the QD and the QEs, so the upgrade must maintain this invariant. |
| Further, the upgrade must do as much as it can in parallel to ensure |
| the cluster is upgraded with as little downtime as possible. To |
| satisfy both requirements, the upgrade process is split between QD and |
| QE. The QD is first upgraded, and the schema dumped and restored. |
| The new data directory created as part of the upgrade is transferred |
| to all QEs, with the local QE configuration retained. The QE upgrade |
| can then skip the schema dump/reload part and only transfer the files. |
| |
| |
| Extensions to upstream pg_upgrade |
| --------------------------------- |
| |
| Functionality for restoring the append-only catalog relations is in |
| aotable.c. |
| |
| Upgrading from 5.0 to 6.0 |
| ------------------------- |
| |
| The name_pattern_ops and pattern equality operators were removed in |
| 6.0 and text equality is treated like bitwise equality. This will |
| cause internal sort order in bpchar_pattern_ops which is now |
| trailing-blank-insensitive. This means that indexes must be dropped |
| and recreated as part of the upgrade to ensure correct results. |
| |
| |
| Testing Apache Cloudberry pg_upgrade |
| ---------------------------- |
| |
| There is a testrunner script in contrib/pg_upgrade which runs a full |
| upgrade of the gpdemo cluster and it's current contents. The idea is |
| that it can be used to test upgrades by running as the final step of |
| ICW. The test can be invoked by "make installcheck" in the |
| contrib/pg_upgrade directory. |
| |
| The testscript also has a "smoketest" option which only tests |
| upgrading a single node. This is *not* a test of an actual Cloudberry |
| Database cluster upgrade, but a test of the Oid synchronization |
| required to handle a node upgrade. The intention is for this to be a |
| quick test that all objects are handled by pg_upgrade; upgrade testing |
| still requires the full run across all segments/mirrors. The smoketest |
| can be invoked by "make check" in the contrib/pg_upgrade directory. |
| GPDB Merge notice: |
| ------------------ |
| We added the gp_fatal_log() function to replace some of the pg_fatal() calls in |
| check.c and check_gp.c to support --continue-check-on-fatal, which allows |
| running through all the checks even if we get a fatal failure. We need to |
| update any newly added checks to ensure they work with the flag by replacing |
| pg_fatal function calls with gp_fatal_log calls. |
| |
| Note: we cannot skip few checks despite using the --continue-check-on-fatal |
| flag. See gp_fatal_log() for details. |
| |
| For --skip-target-check flag while merging upstream, if we have new checks for |
| new/target cluster, they should be skipped when the flag is enabled. See |
| is_skip_target_check(). There might be new pre-check steps for the target |
| cluster before we check_new_cluster(), such as get_sock_dir(), which must be |
| guarded. |
