Changelog
All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
v1.1.0
This release adds optional pgBouncer connection pooling and support for an externally managed Patroni configuration store. Both are opt-in: an inventory that sets neither renders exactly the configuration v1.0.0 did, so existing clusters upgrade in place.
Added
- New
patroni_dcsparameter selects the distributed configuration store Patroni uses and passes arbitrary connection settings to it, allowing an externally managed store such as Consul or ZooKeeper. Accepted types areetcd3,etcd,consul,zookeeper, andexhibitor. Patroni'skubernetestype is not accepted, because it discovers the API server from a pod environment or a kubeconfig rather than from a configured endpoint, and this collection installs Patroni on ordinary hosts. (EE-33) install_patroninow installs the Patroni client library that matches the configuredpatroni_dcstype. (EE-33)- Ultra-HA end-to-end test now covers an externally managed Consul store in addition to the default etcd cluster. (EE-33)
- New
patroni_namespaceparameter sets the key prefix Patroni uses within the distributed configuration store, so a store shared by more than one zone can give each zone its own prefix. Defaults to/db/, which matches the previous hardcoded value. (EE-33) - New
patroni_scopeparameter sets the cluster name Patroni uses within the distributed configuration store, so a store shared by more than one zone can give each zone its own scope instead of its own namespace prefix, which some stores make load-bearing. Defaults to<pg_version>-<cluster_name>, which matches the previous hardcoded value, and thepatronictlinvocations insetup_patroniandsetup_backrestnow name the cluster with it. (EE-33) init_servernow asserts thatpatroni_dcs.parametersis present whenpatroni_dcs.typenames a store the collection does not deploy, so a missing key fails before bootstrapping rather than while Patroni is configured. (EE-33)role_configgained apg_feature_checkstask file that asks the installed PostgreSQL binary which configuration parameters it recognizes.setup_postgresandsetup_patroniinclude it so configuration can be gated on parameters that only some releases carry. (EE-34)- New
install_pgbouncerandsetup_pgbouncerroles deploy a pgBouncer connection pooler on pgEdge nodes, giving each node a second endpoint onpgbouncer_portbeside PostgreSQL's own. Pooling is opt-in through the newpgbouncer_enabledparameter, which is cluster-wide the wayis_ha_clusteris: set it on thepgedgegroup and every node pools, leave it unset and the cluster renders exactly the configuration it did before.init_serverrejects an inventory whose pgEdge nodes disagree, because a zone that pooled only some of its nodes would lose its pooled endpoint on the first failover to one of the others. - Pooled connections authenticate through pgBouncer's
auth_queryrather than a maintained password list.setup_postgrescreates a powerlesspgbouncer_authrole and aSECURITY DEFINERlookup wherever the cluster pools, so every PostgreSQL role works through the pooled endpoint, including roles created after deployment, and a rotated password takes effect immediately. The lookup filters on the role'sVALID UNTIL, so an expired password is refused at the pooler instead of being accepted there and failing the backend login. Onlypgbouncer_auth_passwordis written to disk, andinit_serverrefuses to deploy a pooled cluster while it is still the default. - The pooler enforces its own client authentication rules, rendered into
/etc/pgbouncer/pg_hba.conffrom the same variables that drive the PostgreSQL rules.custom_hba_rulesadmits a client to both endpoints; the newpgbouncer_hba_rulesadmits it to the pooled endpoint alone.init_servervalidates both against thepg_hbasubset pgBouncer can parse, since it skips a line it cannot parse rather than refusing to start.pgbouncer_authreaches the pooler over its unix socket only: the rendered rules reject it on every address, ahead of the loopback and proxy rules that name every role, so the one account that can read stored verifiers is not also a network login. - The pooled endpoint serves TLS from the same certificate PostgreSQL
presents, staged from the controller rather than read out of
PGDATAso an HA replica does not race Patroni's clone.pgbouncer_client_tls_sslmodedefaults toallow, which accepts exactly what the direct endpoint accepts today. - New
pooler_portparameter fronts the poolers from the proxy layer, the wayproxy_portfrontspg_port. Where the cluster pools,setup_haproxyemits apg-poolerlistener on it carrying every node of the zone — the same servers as the direct listener, differing only in the port — health-checked against Patroni's REST API like the direct listeners, and sizes the global connection ceiling to cover it. The existing listeners are unchanged, so Spock replication never routes through a pooler.init_serverasserts the ports a host actually binds do not collide: the two proxy-layer ports against each other on any HAProxy node, and againstpg_portandpgbouncer_portonly where HAProxy shares a host with a pgEdge node, since that is the only topology where all four bind the same address. - New documentation for pooling: role pages for both new roles, a Pooling
Configuration reference, a pgBouncer troubleshooting page, the pooled
listener and the port model in Proxy Configuration,
pgbouncer_enabledin Inventory Structure, and an opt-in walkthrough in both tutorials.
Fixed
- Spock replication no longer breaks on PostgreSQL releases carrying the fix
for CVE-2026-6471, which refuse to load an output plugin that
output_plugin_librariesdoes not name. Both simple and Ultra-HA clusters now set the parameter, and only on releases that recognize it, since earlier releases refuse to start when it appears. (EE-34) init_servernow rejects an inventory that givesproxy_portthe same value aspg_porton a host that runs HAProxy alongside a pgEdge node, a topology the proxy documentation supports. HAProxy could not bind its listener there, which surfaced as a service failure well after the playbook had configured the node rather than as a validation error before it started.init_servervalidations that read a host's own variables now run on the hosts that carry them rather thanrun_once. Ansible runs arun_oncetask on the first host of the play and evaluates its condition only there, so on an inventory that ordered a proxy, backup or client host first, these checks read variables that host did not have — reporting a failure against the wrong host, or skipping silently and reporting nothing at all. Affected the default password check,pgbouncer_auth_password, the pooler's TLS mode and HBA rule checks, the per-zone node check, the supported-OS check and both DCS checks. The two assertions that read the inventory as a whole, and so resolve the same from any host, still run once.patroni_config_fileandpatroni_tls_dirare now recognized by all roles.- HA failover example in the usage guide now passes the Patroni scope the collection actually configures, which has included the PostgreSQL version since v1.0.0.
- Patroni replication user now connects to all databases for logical slot creation.
backup_repo_cipherdefault is now properly deterministic.- PostgreSQL contrib package is now installed explicitly on RHEL systems where it may be missing.
v1.0.0
This release is a major overhaul that revises every role in the collection.
It replaces the pgEdge CLI component with standard RPM and DEB packages from
the pgEdge and PGDG package repositories. All local-user installation
semantics are removed in favor of the standard postgres OS user and
system-conventional paths.
This is a breaking change. Clusters deployed with v0.1 or v0.2 cannot be upgraded in place. Re-provision all nodes from scratch before deploying to production.
Added
- New
install_reposrole installs the pgEdge and PGDG package repositories on each node before any software installation takes place. Add this role to all playbooks immediately afterinit_server. (EE-22) custom_hba_rulesparameter accepts a list of custompg_hba.confrules to append to the default rule set.backup_userandbackup_passwordparameters define a dedicated PostgreSQL user withpg_checkpointprivileges for backup operations.backup_repo_userparameter specifies the OS user that owns the PgBackRest repository in SSH backup mode.proxy_portparameter allows HAProxy to run on a pgEdge node by setting a proxy port separate from the PostgreSQL listen port.pgedge_userandpgedge_passwordparameters define the internal user used for node-to-node Spock connections.tls_validity_daysparameter controls the validity period for generated TLS certificates.
Fixed
- Patroni cluster initialization now waits for the primary to become available before performing administrative operations.
Changed
- Switched to pgEdge distro-specific package repositories;
repo_namedefault changed fromdownloadtorelease. (EE-22) - PostgreSQL, etcd, Patroni, and PgBackRest are now installed from system
packages rather than downloaded locally. The
install_baseandcluster_pathparameters are removed. - PostgreSQL now runs as the system
postgresuser. Data and configuration directories follow OS conventions:/var/lib/postgresql/VERSION/mainand/etc/postgresql/VERSION/CLUSTERon Debian;/var/lib/pgsql/VERSION/dataon RHEL. pg_home,pg_path,pg_data, andpg_config_dirare now computed frompg_versionand OS family. Thecluster_path-based path variables are removed.- Default
pg_hba.confrules now use a least-privilege model and only include entries for known user and database combinations. Custom rules can be added viacustom_hba_rules. init_servernow disablesRemoveIPCin systemd-logind, creates thepostgresOS user on nodes that require SSH backup access, and validates configuration before any other tasks run.- Updated default etcd version to 3.6.5.
- Spock extension is now always installed at the latest available version.
- Several parameters previously scattered across role defaults are now
centralized in the
role_configrole.
v0.2.0
Added
- Ability to specify additional HAProxy listeners for specific Patroni node types. (EE-9)
- Restart Postgres in setup_patroni role to activate config changes. (EE-13)
Changed
- Switched to upstream etcd and Patroni releases. (EE-15)
- Set default Spock version to 5.0.
- Set default Postgres version to 17.
v0.1.0
The "real" release.
Added
- Ability to specify multiple database names for subscription management. (EE-12)
- Ability to specify exception_behaviour as parameter. (EE-7)
- Can now specify proxy_node to override automatic haproxy zone target. (EE-3)
Fixed
- Additional retries to circumvent Debian package management locks. (#15)
- replication_user and replication_password actually work now. (EE-14)
- Strip comments from ssh_keyscan output to avoid syntax errors. (#13)
- DDL replication now activates spock.allow_ddl_from_functions. (EE-10)
- Added lock_timeout parameter to all package tasks to avoid failures. (EE-8)
- Added explicit dependency to ansible.utils.ipaddr filter. (EE-1)
Changed
- Renamed db_name role parameter to db_names to list multiple database names. (EE-12)
v0.0.2
Added
- Initial creation and setup roles for pgEdge cluster resources.