diff options
Diffstat (limited to 'configdefinitions/src/vespa/vespa.config.content.fleetcontroller.def')
| -rw-r--r-- | configdefinitions/src/vespa/vespa.config.content.fleetcontroller.def | 183 |
1 files changed, 183 insertions, 0 deletions
diff --git a/configdefinitions/src/vespa/vespa.config.content.fleetcontroller.def b/configdefinitions/src/vespa/vespa.config.content.fleetcontroller.def new file mode 100644 index 00000000000..96b43a15c5e --- /dev/null +++ b/configdefinitions/src/vespa/vespa.config.content.fleetcontroller.def @@ -0,0 +1,183 @@ +# Copyright 2017 Yahoo Holdings. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root. +namespace=vespa.config.content + +## Name of VDS cluster +cluster_name string restart + +## The fleetcontroller index. Each fleetcontroller should have a unique index +## which can be used to identify them. +index int restart + +## Number of fleet controllers. If more than one, the fleetcontroller needs to +## do master election in order to know which one is active. It then needs to +## know how many exist. +fleet_controller_count int default=1 + +## The number of seconds we can attempt to talk to zookeeper before our session +## time out and we lose our connection. +zookeeper_session_timeout double default=30.0 + +## When a master candidate see the master disappearing from ZooKeeper, it wants +## to take over as master. But, before taking over, the master should be aware +## that it has lost its zookeeper connection, so it will stop reacting as +## master. Suggests setting this to 2-3 times the zookeeper session timeout. +## (Set in number of seconds) +master_zookeeper_cooldown_period double default=60.0 + +## Sets how many fleetcontrollers will gather state. A fleetcontroller +## gathering state can take over quicker should ## the master fail. +## If set to 1, only master will gather state. If set higher, others will +## also do so, prioritizing those fleetcontrollers likely to be the ones to +## take over if the master fails. +state_gather_count int default=2 + +## Location of ZooKeeper servers +zookeeper_server string restart + +## RPC Port used by fleetcontroller +rpc_port int default=6500 restart + +## Port where fleetcontroller listens for HTTP status request +http_port int default=0 restart + +## Maximum number of milliseconds a storage will be automatically reported in +## maintenance due to node recently being availble. (If 0, no time at all) +storage_transition_time int default=30000 + +## Maximum number of milliseconds a distributor will be automatically reported +## in maintenance due to node recently being availble. (If 0, no time at all) +## +## Currently default is 0.. Should probably be more then we know it is working +## correctly +distributor_transition_time int default=0 + +## Maximum number of milliseconds allowed between progress increase during +## initialization. If no progress have been made during this time period, the +## node will be considered down. +## +## Currently disabled as 5.0 nodes gets load while initializing which may be +## higher pri than initializing, so nodes can stay in init for a long time. +init_progress_time int default=0 + +## Minimum time (in ms) between system state updates. To limit updates in a +## system where a lot is happening at the same time, this value will make sure +## we dont change the state too often. +min_time_between_new_systemstates int default=10000 + +## Sets how many milliseconds to wait between each state poll for old nodes +## requiring state polling. (4.1 or older) +state_polling_frequency int default=5000 + +## The maximum amount of premature crashes a node is allowed to have in a row +## before the fleetcontroller disables that node. +max_premature_crashes int default=100000 + +## If a node has been down or up this many milliseconds, clear the premature +## crash count of a node and consider the node as stable +stable_state_time_period int default=7200000 + +## The maximum number of events to keep in the event log +event_log_max_size int default=1024 + +## The maximum number of node events to keep in the node event log per node +event_node_log_max_size int default=1024 + +## The total number of distributor nodes that can exist. If 0, we dont know and +## will use the highest distributor index number we have ever seen + 1. +total_distributor_count int default=0 + +## The total number of storage nodes that can exist. If 0, we dont know and +## will use the highest storage index number we have ever seen + 1. +total_storage_count int default=0 + +## The minimum number of distributor nodes that should be up for the cluster +## state to be up. (Retired nodes counts as up in this case) +min_distributors_up_count int default=1 + +## The minimum number of storage nodes that should be up for the cluster state +## to be up (Retired nodes counts as up in this case) +min_storage_up_count int default=1 + +## The minimum ratio of known distributor nodes that should be up (or retired) +## for the cluster state to stay up. +min_distributor_up_ratio double default=0.01 + +## The minimum ratio of known storage nodes that should be up (or retired) for +## the cluster state to stay up. +min_storage_up_ratio double default=0.01 + +## Seconds to sleep after doing a work cycle where we did no work. Some +## events do not interrupt the sleeping, such as slobrok changes, so shouldn't +## set this too high +cycle_wait_time double default=0.1 + +## Minimum time to pass in seconds before broadcasting our first systemstate as +## a new fleetcontroller. (Will broadcast earlier than this if we have gathered +## state from all before this). To prevent disturbance when taking over as +## fleetcontroller, give nodes a bit of time to answer so we dont temporarily +## report nodes as down. +min_time_before_first_system_state_broadcast double default=5.0 + +## Request timeout of node state requests. Keeping a high timeout allows us to +## always have a pending operation with very low cost. Keeping a low timeout is +## good to detect issues like packet loss. The default tries to balance the two +## by not resending too often, but detecting packet loss within a minute at +## least. If we can guarantee RPC layer to fail on packet loss within +## reasonable time we should increase this default. +get_node_state_request_timeout double default=120.0 + +## If a node is out of slobrok longer than this time period, assume the node +## is down, even if we have a pending node state request to it. Slobrok does +## a bit more keep alive checking than fleetcontroller, so it is possible that +## the node disappears from slobrok while it still looks ok in fleetcontroller. +max_slobrok_disconnect_grace_period double default=60.0 + +## Whether to show system states that have never been sent to storage nodes in +## the event log. +show_local_systemstates_in_event_log bool default=true + +## The ideal number of distribution bits this system should have +ideal_distribution_bits int default=16 + +## Minimum ratio of nodes that have to be available (i.e. not Down) in any +## hierarchic content cluster group. If a higher ratio than this is Down at +## any point, the remaning nodes in the group will be automatically marked +## as down. Group nodes will automatically be taken back up as soon as node +## availability has been restored above the given threshold. +## Default is 0, i.e. functionality is for all intents and purposes disabled. +min_node_ratio_per_group double default=0.0 + +## If a cluster controller task has a dependency on a given cluster state +## version being published and ACKed by the cluster, it will be put on a wait +## queue while holding up the container thread associated with the task. +## This config specifies the maximum time a task can be held in this queue +## before being automatically failed out, if a version has not been ACKed +## within this duration. +max_deferred_task_version_wait_time_sec double default=30.0 + +## Whether or not the content cluster the controller has responsibility for +## contains any document types that are tagged as global. If this is true, +## global document-specific behavior is enabled that marks nodes down in the +## default space if they have merges pending in the global bucket space. +cluster_has_global_document_types bool default=false + +## The minimum merge completion ratio of buckets in a bucket space before it is considered complete. +## +## Bucket merges are considered complete when: +## ((buckets_total - buckets_pending) / buckets_total)) >= min_merge_completion_ratio +min_merge_completion_ratio double default=1.0 + +## If enabled, cluster state transitions are performed as two distinct phases: +## +## 1) state bundle propagation and bucket info gathering phase +## 2) state activation phase, which is not performed until all nodes have completed phase 1 +## +## This is to enable read-only operations to pass through the system during phase 1 +## while nodes await phase 2. If this feature is disabled, nodes will implicitly do +## phase 2 as part of phase 1 at their own leisure, which means that actual state +## activation may happen at wildly different times throughout the cluster. The 2 phase +## transition logic aims to minimize the window of time where active states diverge. +enable_two_phase_cluster_state_transitions bool default=false + +## Deprecated - not used +determine_buckets_from_bucket_space_metric bool default=true |