Change Details

This property is dangerous in that it prevents Pacemaker from gaining an accurate view of resource state, without preventing other actions as (for example) maintenance mode would do. If the property is set to false, we log an info-level message that notes that it's dangerous: ``` set_config_flag(scheduler, PCMK_OPT_ENABLE_STARTUP_PROBES, pcmk__sched_probe_resources); if (!pcmk__is_set(scheduler->flags, pcmk__sched_probe_resources)) { crm_info("Startup probes: disabled (dangerous)"); } ``` However, there is no such warning in the option documentation. The option is not even flagged as advanced use only. ``` { PCMK_OPT_ENABLE_STARTUP_PROBES, NULL, PCMK_VALUE_BOOLEAN, NULL, PCMK_VALUE_TRUE, pcmk__valid_boolean, pcmk__opt_schedulerd, N_("Whether the cluster should check for active resources during " "start-up"), NULL, }, ``` At the very least, we should add the `pcmk__opt_advanced` flag. However, this property seems like a bad one to expose. At least `resource-discovery=never` is configured in a location constraint, so that it affects only named resources (or resources matching a regex). The common use case there is to prevent failures on nodes that will never be allowed to run a given resource and may not even have the software installed. This option was introduced by commit b20fd762, and the expressed motivation was that the "calculation [of startup probes] is a major bottleneck for very large clusters." I doubt that's a major concern on modern hardware, and even if it is, Pacemaker still should have a correct view of resource state.

This property is dangerous in that it prevents Pacemaker from gaining an accurate view of resource state, without preventing other actions as (for example) maintenance mode would do. If the property is set to false, we log an info-level message that notes that it's dangerous: ``` set_config_flag(scheduler, PCMK_OPT_ENABLE_STARTUP_PROBES, pcmk__sched_probe_resources); if (!pcmk__is_set(scheduler->flags, pcmk__sched_probe_resources)) { crm_info("Startup probes: disabled (dangerous)"); } ``` However, there is no such warning in the option documentation. The option is not even flagged as advanced use only. ``` { PCMK_OPT_ENABLE_STARTUP_PROBES, NULL, PCMK_VALUE_BOOLEAN, NULL, PCMK_VALUE_TRUE, pcmk__valid_boolean, pcmk__opt_schedulerd, N_("Whether the cluster should check for active resources during " "start-up"), NULL, }, ``` At the very least, we should add the `pcmk__opt_advanced` flag. However, this property seems like a bad one to expose. At least `resource-discovery=never` is configured in a location constraint, so that it affects only named resources (or resources matching a regex). The common use case there is to prevent failures on nodes that will never be allowed to run a given resource and may not even have the software installed. This option was introduced by commit b20fd762, and the expressed motivation was that the "calculation [of startup probes] is a major bottleneck for very large clusters." I doubt that's a major concern on modern hardware, and even if it is, Pacemaker still should have a correct view of resource state. Also, by inspection (have not tested) of `pcmk__probe_rsc_on_node()`, it looks like setting this property to false actually disables all probes, not just probes during startup.

This property is dangerous in that it prevents Pacemaker from gaining an accurate view of resource state, without preventing other actions as (for example) maintenance mode would do. If the property is set to false, we log an info-level message that notes that it's dangerous: ``` set_config_flag(scheduler, PCMK_OPT_ENABLE_STARTUP_PROBES, pcmk__sched_probe_resources); if (!pcmk__is_set(scheduler->flags, pcmk__sched_probe_resources)) { crm_info("Startup probes: disabled (dangerous)"); } ``` However, there is no such warning in the option documentation. The option is not even flagged as advanced use only. ``` { PCMK_OPT_ENABLE_STARTUP_PROBES, NULL, PCMK_VALUE_BOOLEAN, NULL, PCMK_VALUE_TRUE, pcmk__valid_boolean, pcmk__opt_schedulerd, N_("Whether the cluster should check for active resources during " "start-up"), NULL, }, ``` At the very least, we should add the `pcmk__opt_advanced` flag. However, this property seems like a bad one to expose. At least `resource-discovery=never` is configured in a location constraint, so that it affects only named resources (or resources matching a regex). The common use case there is to prevent failures on nodes that will never be allowed to run a given resource and may not even have the software installed. This option was introduced by commit b20fd762, and the expressed motivation was that the "calculation [of startup probes] is a major bottleneck for very large clusters." I doubt that's a major concern on modern hardware, and even if it is, Pacemaker still should have a correct view of resource state. Also, by inspection (have not tested) of `pcmk__probe_rsc_on_node()`, it looks like setting this property to false actually disables all probes, not just probes during startup.