Monitor Server¶
Monitor Server is component responsible for providing infrastructure necessary for multiple redundant component execution. Additionally, Monitor Server provides centralized registry of running components in single system which can be used for service discovery.
Warning
Monitor Server functionality is defined under assumptions:
components running as part of single system are trusted and implemented according to specifications
communication channels between components on single computing node are reliable
Redundancy mechanisms provided by Monitor Server do not provide reliability in case of possible component malfunctions due to implementation errors.
Running¶
By installing Monitor Server from hat-monitor package, executable hat-monitor becomes available and can be used for starting this component.
Usage: __main__.py [OPTIONS] Main entry point Options: --conf PATH configuration defined by hat://monitor/main.yaml# (default $XDG_CONFIG_HOME/hat/monitor.yaml) --ui-path PATH Override web ui directory path (development argument) --help Show this message and exit.
Additionally, hat-monitor package provides implementation of library which
can be used as basis for communication between components and Monitor Server.
This library is available in hat.monitor.client module.
Communication model¶
Monitor Server provides n-node redundancy architecture based on chatter protocol (structure of communication messages is defined in HatMonitor package). It is based on server-client communication between components and monitor server. There also exists horizontal peer communication between multiple monitor servers which enables forming of single system based on multiple distributed computing nodes. It is assumed that, for each system, single computing node runs single instance of Monitor Server and unlimited number of components that connect to local Monitor Server. Each node’s Monitor Server establishes communication with other nodes’ Monitor Server through single master Monitor Server. Hierarchy of Monitor Servers which can perform functionality of master is configured statically with configuration options available to each Monitor Server on its startup.
Entities participating in communication between components and local monitor server:
Client
Component proclaiming its existence to the server, discovering other components and participating in redundancy algorithm.
Server
Local monitor server providing global components state to all local clients and providing user interface.
Entities participating in ‘horizontal’ communication between Monitor Servers:
Master
Monitor Server responsible for execution of the redundancy algorithm and notifying other Monitor Servers about the current global state of the entire system.
Slave
Monitor Server responsible for notifying master about its current local state and delegating master’s notifications to its server entity.
Monitor Server uses two different independent listening sockets for client-server and master-slave communication.
If we represent components with Cn and Monitor Servers with Mn, where master hierarchy of M1 > … > Mn is presumed, an example of a single system monitor communication can be viewed as:
graph { bgcolor=transparent; layout=neato; node [fontname="Arial"]; { node [shape=box]; M1; M2; M3; M4; } { node [shape=oval]; C1; C2; C3; C4; C5; C6; C7; C8; C9; C10; C11; C12; } { edge [dir=forward, style=bold, len=2]; M2 -- M1; M3 -- M1; M4 -- M1; } { edge [dir=forward, style="bold,dashed", len=2]; M3 -- M2; M4 -- M2; M4 -- M3; } { edge [dir=both, arrowsize=0.5]; M1 -- C1; M1 -- C2; M1 -- C3; M2 -- C4; M2 -- C5; M2 -- C6; M3 -- C7; M3 -- C8; M3 -- C9; M4 -- C10; M4 -- C11; M4 -- C12; } }
Component information¶
Component information is basic structure of properties that describe each component included in system. It is initially created on local Monitor Server and later updated by master Monitor Server. Collection of all components information associated with clients connected to local Monitor Server and calculated by local Monitor Server is called local state. Collection of all components information in single system calculated by master Monitor server is called global state. Each Monitor Server provides global state to its local clients.
Properties included in a component information:
cid
Component id assigned to client by its local Monitor Server.
mid
Monitor id identifying local Monitor Server (assigned to local Monitor Server by master). Value
0indicates Monitor Server which is master or is not connected to remote master.name
User provided identifier of component. This entry is used only for UI presentation purposes and logging but it is recommended to use unique identifiers for each component instance. This property is assigned by client.
group
String identifier by which components are grouped while blessing calculation algorithm is applied (see Blessing algorithm). This property is assigned by client.
address
Component’s address which other components use to connect to the component. This property is optional and is assigned by client.
rank
Component’s rank - used by Blessing algorithm. This property is initially assigned by local Monitor Server but can later be changed only by local Monitor Server.
blessing
Optional number used as unique token assigned and changed by master Monitor Server (see Component lifetime).
ready
Optional number used as unique token assigned and changed by client (see Component lifetime).
Master slave communication¶
Horizontal communication between Monitor Servers is hierarchically ordered.
Each Monitor Server knows its superiors’ addresses. If M1 > M2 > M3,
then M1 doesn’t know any other monitor address; M2 knows the address
of M1; M3 knows addresses of M1 and M2 in that order.
Each Monitor Server’s configuration contains zero or more other Monitor Server addresses. These other servers are “superior” to the monitor server. A monitor server will always try to maintain an active connection with exactly one of its superiors. The addresses list is ordered by priority meaning that if the Monitor Server isn’t connected to a superior, it tries to connect to the first monitor server in the list. If the connection fails, it tries the second one and so on. If it can’t connect to any of its superiors, it can proclaim itself as master. The connecting to master process should continue even if the Monitor Server is master or it is not connected to its first superior. The connecting loop should be executed at least 2 or 3 times before the timeout can be used. The timeout should be kept small.
Once a slave Monitor Server connects to the Master Monitor server it sends its local state to the master and keeps notifying the master about any change in its local state while the connection is active. The master gathers all local states and generates its global state which it then transmits to all its slaves and keeps notifying them of any change. Global state contains all component informations received from local states except for those where component’s name or group are not set. Master also identifies each Monitor Server with unique monitor identifier (mid) which is provided to slave together with global state. It is important to note that only master Monitor Server calculates blessing token state for each component.
Every Monitor Server is responsible for starting master listening socket immediately after startup. While Monitor Server isn’t operating in master mode, all connections made to master listening socket will be closed immediately after their establishment - this behavior will indicate to connecting Monitor Server that its superior is not currently master.
Messages used in master slave communications are:
Message
Conversation
Direction
First
Last
Token
MsgSlave
T
T
T
s > m
MsgMaster
T
T
T
m > s
where s > m represents slave to master communication and m > s represents master to slave communication. When new connection is established, master should immediately associate new mid with connection and send MsgMaster. After slave receives this initial MsgMaster, it should send MsgSlave with local state updated with newly received mid. Each communicating entity (master or slave) should send new state message (MsgMaster or MsgSlave) if any data obtained from MsgSlave or MsgMaster changes. Sending of MsgMaster and MsgSlave should be implemented independent of receiving messages from associated entity. Implementation of master should not be dependent on receiving initial MsgSlave and should continue sending MsgMaster on every state change even if no MsgSlave is received.
Server client communication¶
Vertical communication between client and server enables bidirectional asynchronous exchange of component information data. Client is responsible for providing name, group, address and ready properties initially and on every change. Server provides global state to each connected client and each client’s component id (cid) and monitor id (mid). If any part of state available to server changes (including token changes), server sends updated state to all clients. Client can also request change for information provided to server at any time.
Messages used in server client communications are:
Message
Conversation
Direction
First
Last
Token
MsgClient
T
T
T
c > s
MsgServer
T
T
T
s > c
where c > s represents client to server communication and s > c represents server to client communication. When new connection is established, each communicating entity (server or client) server immediately sends initial state message (MsgServer or MsgClient) and should send new state messages when any data obtained from MsgServer or MsgClient changes. Sending of MsgServer and MsgClient should be implemented independent of receiving messages from associated entity. Implementation of server should not be dependent on receiving initial MsgClient and should continue sending MsgServer on every state change even if no MsgClient is received.
Component lifetime¶
For most components, connection to local Monitor Server is mandatory for providing implemented functionality. Because of this, component usually connects to local Monitor Server during startup and preserves this active connection during entire component run lifetime. If this connection is closed for any reason, process also terminates. This behavior is not mandated.
Components which connect to Monitor Server participate in redundancy supervised by master Monitor Server. Redundancy utilizes two tokens:
blessing token
This token is controlled exclusively by master Monitor Server. If connection to master is not established, token’s value is not set.
ready token
This token is controlled exclusively by client. It will match blessing value only if component is ready to provide its primary functionality. At any time, if component stops providing primary functionality, it should revoke this token.
Each component determines if it should provide primary functionality based on global state provided by local Monitor Server. If client’s component information contains blessing and ready token with same value, component can provide primary functionality. If, at any time, these values do not match, component should stop its usual activity which is indicated by client’s revoking of ready flag.
Blessing algorithm¶
Blessing algorithm determines value of each component’s blessing token. This calculation is performed on master Monitor Server and should be executed each time any part of global state changes. This calculation should be integrated part of state change and thus provide global state consistency.
Monitor Server implements multiple algorithms for calculating value of blessing token. Each component group can have different associated blessing algorithm and all groups that don’t have associated blessing algorithm use default algorithm. Group’s associated algorithms and default algorithm are provided to Monitor Server as configuration parameters during its startup.
Calculation of blessing token values is based only on previous global state and new changes that triggered execution of blessing algorithm.
Currently supported algorithms:
BLESS_ALL
This simple algorithm provides continuous blessing to all components in associated group. Blessing is never revoked.
BLESS_ONE
In each group with this algorithm associated, there can be only one highlander and component with issued blessing token. For determining which component can receive blessing token, component’s rank is used. Components with mathematically lower rank value have higher priority in obtaining blessing token. If there exist more than one component with highest priority than one with already set blessing token is chosen. If neither of component have already set blessing token, than one of components with lowest mid value is chosen. Once component which can obtain blessing token is chosen, if chosen component doesn’t already have blessing token, master revokes previously issued blessing token from other component in the same group and waits for all components in the same group to revoke theirs ready tokens. Only once all other components revoke their ready tokens, master issues new blessing token to chosen component.
Components rank¶
Association of component’s rank is responsibility of component’s local Monitor Server for all of it’s local components. Monitor Server should associate same rank as was last rank value associated with previously active client connection with same name and group values as newly established connection. If such previously active connection does not exist, default rank value, as specified by Monitor Server’s configuration, is applied. After initial rank value is associated with client and its ComponentInfo, local Monitor Servers can later change rank’s value. These changes should be cached by local Monitor Servers in case connection to component is lost and same component tries to establish new connection. This cache is maintained for duration of single Monitor Server process execution and is not persisted between different Monitor Server processes.
User interface¶
As secondary functionality, Monitor Server provides web-based user interface for monitoring global components state and controlling component’s rank. Implementation of this functionality is split into server-side web backend and web frontend exchanging communication messages based on juggler communication protocol.
Backend to frontend communication¶
Backend provides to frontends all information that is made available by server to clients. When this information changes, all frontends are notified of this change. Current state of mid and all components is set and continuously updated as part of server’s juggler local data.
After new Juggler connection is established, backend will immediately set juggler local data defined by JSON schema:
"$schema": "http://json-schema.org/schema#"
type: object
required:
- mid
- local_components
- global_components
properties:
mid:
type: integer
local_components:
type: array
items:
type: object
required:
- cid
- name
- group
- address
- rank
properties:
cid:
type: integer
name:
type:
- string
- "null"
group:
type:
- string
- "null"
address:
type:
- string
- "null"
rank:
type: integer
global_components:
type: array
items:
type: object
required:
- cid
- mid
- name
- group
- address
- rank
- blessing
- ready
properties:
cid:
type: integer
mid:
type: integer
name:
type:
- string
- "null"
group:
type:
- string
- "null"
address:
type:
- string
- "null"
rank:
type: integer
blessing:
type:
- integer
- "null"
ready:
type:
- integer
- "null"
Server doesn’t send additional MESSAGE juggler messages.
Frontend to backend communication¶
This communication is used primary for enabling user control of components’ ranks. At any time, frontend can send set_rank message to backend requesting change of rank for any available local component. These massages are transmitted as juggler’s MESSAGE messages defined by JSON schema:
"$schema": "http://json-schema.org/schema#"
oneOf:
- "$ref": "#/definitions/set_rank"
definitions:
set_rank:
type: object
required:
- type
- payload
properties:
type:
enum:
- set_rank
payload:
type: object
required:
- cid
- rank
properties:
cid:
type: integer
rank:
type: integer
Client’s juggler local data isn’t changed during communication with server (it remains null).
Future features¶
Todo
configurable master retry count and timeouts
optional connection to monitor/event server
mapping of current status to events
listening for control events