ceph/src/pybind/mgr/cephadm/migrations.py

   1 import logging
   2 from typing import TYPE_CHECKING, Iterator
   3
   4 from ceph.deployment.service_spec import PlacementSpec, ServiceSpec, HostPlacementSpec
   5 from cephadm.schedule import HostAssignment
   6
   7 from orchestrator import OrchestratorError
   8
   9 if TYPE_CHECKING:
  10     from .module import CephadmOrchestrator
  11
  12 LAST_MIGRATION = 2
  13
  14 logger = logging.getLogger(__name__)
  15
  16
  17 class Migrations:
  18     def __init__(self, mgr: "CephadmOrchestrator"):
  19         self.mgr = mgr
  20
  21         # Why having a global counter, instead of spec versions?
  22         #
  23         # for the first migration:
  24         # The specs don't change in (this) migration. but the scheduler here.
  25         # Adding the version to the specs at this time just felt wrong to me.
  26         #
  27         # And the specs are only another part of cephadm which needs potential upgrades.
  28         # We have the cache, the inventory, the config store, the upgrade (imagine changing the
  29         # upgrade code, while an old upgrade is still in progress), naming of daemons,
  30         # fs-layout of the daemons, etc.
  31         if self.mgr.migration_current is None:
  32             self.set(0)
  33
  34         # for some migrations, we don't need to do anything except for
  35         # setting migration_current = 1.
  36         # let's try to shortcut things here.
  37         self.migrate()
  38
  39     def set(self, val):
  40         self.mgr.set_module_option('migration_current', val)
  41         self.mgr.migration_current = val
  42
  43     def is_migration_ongoing(self):
  44         return self.mgr.migration_current != LAST_MIGRATION
  45
  46     def verify_no_migration(self):
  47         if self.is_migration_ongoing():
  48             # this is raised in module.serve()
  49             raise OrchestratorError(
  50                 "cephadm migration still ongoing. Please wait, until the migration is complete.")
  51
  52     def migrate(self):
  53         if self.mgr.migration_current == 0:
  54             if self.migrate_0_1():
  55                 self.set(1)
  56
  57         if self.mgr.migration_current == 1:
  58             if self.migrate_1_2():
  59                 self.set(2)
  60
  61     def migrate_0_1(self) -> bool:
  62         """
  63         Migration 0 -> 1
  64         New scheduler that takes PlacementSpec as the bound and not as recommendation.
  65         I.e. the new scheduler won't suggest any new placements outside of the hosts
  66         specified by label etc.
  67
  68         Which means, we have to make sure, we're not removing any daemons directly after
  69         upgrading to the new scheduler.
  70
  71         There is a potential race here:
  72         1. user updates his spec to remove daemons
  73         2. mgr gets upgraded to new scheduler, before the old scheduler removed the daemon
  74         3. now, we're converting the spec to explicit placement, thus reverting (1.)
  75         I think this is ok.
  76         """
  77
  78         def interesting_specs() -> Iterator[ServiceSpec]:
  79             for s in self.mgr.spec_store.specs.values():
  80                 if s.unmanaged:
  81                     continue
  82                 p = s.placement
  83                 if p is None:
  84                     continue
  85                 if p.count is None:
  86                     continue
  87                 if not p.hosts and not p.host_pattern and not p.label:
  88                     continue
  89                 yield s
  90
  91         def convert_to_explicit(spec: ServiceSpec) -> None:
  92             placements = HostAssignment(
  93                 spec=spec,
  94                 get_hosts_func=self.mgr._get_hosts,
  95                 get_daemons_func=self.mgr.cache.get_daemons_by_service
  96             ).place()
  97
  98             existing_daemons = self.mgr.cache.get_daemons_by_service(spec.service_name())
  99
 100             # We have to migrate, only if the new scheduler would remove daemons
 101             if len(placements) >= len(existing_daemons):
 102                 return
 103
 104             old_hosts = {h.hostname: h for h in spec.placement.hosts}
 105             new_hosts = [
 106                 old_hosts[d.hostname] if d.hostname in old_hosts else HostPlacementSpec(
 107                     hostname=d.hostname, network='', name='')
 108                 for d in existing_daemons
 109             ]
 110
 111             new_placement = PlacementSpec(
 112                 hosts=new_hosts,
 113                 count=spec.placement.count
 114             )
 115
 116             new_spec = ServiceSpec.from_json(spec.to_json())
 117             new_spec.placement = new_placement
 118
 119             logger.info(f"Migrating {spec.one_line_str()} to explicit placement")
 120
 121             self.mgr.spec_store.save(new_spec)
 122
 123         specs = list(interesting_specs())
 124         if not specs:
 125             return True  # nothing to do. shortcut
 126
 127         if not self.mgr.cache.daemon_cache_filled():
 128             logger.info("Unable to migrate yet. Daemon Cache still incomplete.")
 129             return False
 130
 131         for spec in specs:
 132             convert_to_explicit(spec)
 133
 134         return True
 135
 136     def migrate_1_2(self) -> bool:
 137         """
 138         After 15.2.4, we unified some service IDs: MONs, MGRs etc no longer have a service id.
 139         Which means, the service names changed:
 140
 141         mon.foo -> mon
 142         mgr.foo -> mgr
 143
 144         This fixes the data structure consistency
 145         """
 146         bad_specs = {}
 147         for name, spec in self.mgr.spec_store.specs.items():
 148             if name != spec.service_name():
 149                 bad_specs[name] = (spec.service_name(), spec)
 150
 151         for old, (new, old_spec) in bad_specs.items():
 152             if new not in self.mgr.spec_store.specs:
 153                 spec = old_spec
 154             else:
 155                 spec = self.mgr.spec_store.specs[new]
 156             spec.unmanaged = True
 157             self.mgr.spec_store.save(spec)
 158             self.mgr.spec_store.rm(old)
 159
 160         return True