# UniConfig Clustering

# Introduction

UniConfig can be easily deployed in a cluster owing to its stateless architecture and transaction isolation:

  • Stateless architecture - UniConfig nodes in the cluster do not keep any state that needs to be communicated directly to other UniConfig nodes in the cluster. All network-topology configuration and state information is stored inside a PostgreSQL database that must be reachable from all UniConfig nodes in the same zone. All UniConfig nodes share the same database, making the database the single source of truth for the cluster.
  • Transaction isolation - Load-balancing is based on mapping UniConfig transactions to UniConfig nodes in a cluster (transactions are sticky). One UniConfig transaction cannot span multiple UniConfig nodes in a cluster. Southbound sessions used for device management are ephemeral, i.e., they are created when UniConfig needs to access a device on the network (like pushing configuration updates) and are closed as soon as a UniConfig transaction is committed or closed.

There are several advantages to clustered deployment of UniConfig nodes:

  • Horizontal scalability - Increasing the number of units that can process UniConfig transactions in parallel. A single UniConfig node tends to have limited processing and networking resources - this constraint can be mitigated by increasing the number of nodes in the cluster. The more UniConfig nodes in the cluster, the more transactions can be executed in parallel. The number of connected UniConfig nodes in the cluster can also be adjusted at runtime.
  • High availability - A single UniConfig node does not represent a single point of failure. If a UniConfig node crashes, only transactions processed by the corresponding node are cancelled. The application can retry a failed transaction, which will be processed by the next node in the cluster.

There also are a couple limitations to be considered:

  • Parallel execution of transactions is subject to a locking mechanism, whereby two transactions cannot manipulate the same device at the same time.
  • A single transaction is always executed by a single UniConfig node. This means that the scope of a single transaction is limited by the number of devices and their configurations that a single Uniconfig node can handle.

# Deployments

# Single-zone deployment

In a single-zone deployment, all managed network devices are reachable by all UniConfig nodes in the cluster zone. Components of a single-zone deployment and connections between them are illustrated in the diagram below.

Deployment with single zone
Deployment with single zone

Included components:

  • UniConfig controllers - Network controllers that use a common PostgreSQL system for persistence of data, communicate with network devices using NETCONF/GNMi/CLI management protocols and propagate notifications into Kafka topics (UniConfig nodes act only as Kafka producers). UniConfig nodes do not communicate with each other directly, their operations can only be coordinated using data stored in the database.
  • Database storage - PostgreSQL is used for persistence of network-topology configuration, mountpoints settings and selected operational data. The PostgreSQL database can also be deployed in the cluster (outside of scope).
  • Message and notification channels - The Kafka cluster is used to propagate notifications generated by UniConfig itself (e.g., audit and transaction notifications) or from network devices and only propagated by UniConfig controllers.
  • Load-balancers - Load-balancers are used to distribute transactions (HTTP traffic) and SSH sessions from applications to UniConfig nodes. From the point of view of the load-balancer, all UniConfig nodes in the cluster are equal. Currently only a round-robin load-balancing strategy is supported.
  • Managed network devices - Devices that are managed using NETCONF/GNMi/CLI protocols by UniConfig nodes or generate notifications to UniConfig nodes. Sessions between UniConfig nodes and devices are either on-demand/emphemeral (configuration of devices) or long-term (distribution of notifications over streams).
  • HTTP/SSH clients and Kafka consumers - Application layer, such as workflow management systems or end-user systems. The RESTCONF API is exposed using the HTTP protocol, the SSH server exposes UniConfig shell and Kafka brokers allow Kafka consumers to listen to events on subscribed topics.

# Multi-zone deployment

This type of deployment has multiple zones that manage separate sets of devices for the following reasons:

  • Network reachability issues - Groups of devices are reachable, and thus manageable, only from some part of the network (zone) but not from others
  • Logical separation - Different scaling strategies or requirements for different zones.
  • Legal issues - Some devices must be managed separately with split storage because of, for example, regional restrictions.

The following diagrams represents a sample deployment with two zones:

  • Zone 1 contains three UniConfig nodes.
  • Zone 2 contains only two UniConfig nodes.

Multiple zones can share a single Kafka cluster, but database instances must be be split (can be running in a single Postgres server).

Deployment with multiple zones
Deployment with multiple zones

Description of multi-zone areas:

  • Applications - The application layer is responsible for managing mapping between network segments and UniConfig zones. Typically this is achieved by deploying/using an additional inventory database that contains device <-> zone mappings, and based on this information the application decides which zone to use.
  • Isolated zones - A zone contains one or more UniConfig nodes, load-balancers and managed network devices. Clusters in isolated zones do not share information.
  • PostgreSQL databases - It is necessary to use a dedicated database per zone.
  • Kafka cluster - A Kafka cluster can be shared by multiple clusters in different zones, or alternatively there can be a single Kafka cluster per zone. Notifications from different zones can be safely pushed to common topics, as there can be no conflicts between Kafka publishers. However, it is also possible to achieve isolation of published messages in a shared Kafka deployment by setting different topic names in different zones.

# Load-balancer operation

Load-balancers are responsible for allocating UniConfig transactions to one of the UniConfig nodes in the cluster. This is done by forwarding requests without a UniConfig transaction header to one of the UniConfig nodes (using round-robin strategy) and afterwards appending a backed identifier to the create-transaction RPC response in the form of an additional Cookie header ('sticky session' concept). Afterwards, the application is responsible for assuring that all requests belonging to the same transaction contain the same backend identifier.

The following sequence diagram captures the process of creating and using two UniConfig transactions with a focus on load-balancer operation.

Load-balancing UniConfig transactions
Load-balancing UniConfig transactions

  • The first create-transaction RPC is forwarded to the first UniConfig node (applying round-robin strategy), because it does not contain the uniconfig_server_id key in the Cookie header. The response contains both the UniConfig transaction ID (UNICONFIGTXID) and uniconfig_server_id that represents a 'sticky cookie'. The uniconfig_server_id cookie header is appended to the response by the load-balancer.
  • The next request that belongs to the created transaction contains the same UNICONFIGTXID and uniconfig_server_id. The load balancer uses uniconfig_server_id to forward this request to the correct UniConfig node.
  • The last application request again represents the create-transaction RPC. This time the request is forwarded to the next registered UniConfig node in the cluster according to the round-robin strategy.

# Configuration

# UniConfig configuration

All UniConfig nodes in the cluster should be configured using the same parameters. There are several important sections in the config/application.properties file that relate to the clustered environment.

# Database connection settings

This section contains information on how to connect to a PostgreSQL database as well as connection pool settings. It is located under the dbPersistence.connection properties object.

Example with essential settings:

database connection settings
# DB persistence settings

# This property controls the time (in milliseconds) that a client will wait for a connection from the pool.
# If this value is exceeded without a connection becoming available, a {@link SQLException} is thrown.
# The lowest acceptable connection timeout is 250 ms. The default value is 30000 (i.e., 30 seconds)
# This property controls the maximum lifetime of a connection in the pool.
# On a connection-by-connection basis, minor negative attenuation is applied to avoid mass extinction in the pool.
# We strongly recommend setting this value, and it should be several seconds shorter than any database or
# infrastructure imposed connection time limit. A value of 0 indicates no maximum lifetime (infinite lifetime),
# subject to the idleTimeout setting.
# The minimum allowed value is 30000ms (30 seconds). Default: 1800000 (30 minutes).
# This property controls the minimum number of idle connections that HikariCP tries to maintain in the pool.
# If the number of idle connections dips below this value and total connections in the pool are less than maximumPoolSize,
# HikariCP will attempt to add additional connections quickly and efficiently.
# However, for maximum performance and responsiveness to spike demands, we recommend not setting this value
# and instead allowing HikariCP to act as a fixed size connection pool. Default: same as maximumPoolSize
# This property controls the maximum size that the pool is allowed to reach,
# including both idle and in-use connections. Basically this value will determine the maximum number
# of actual connections to the database backend. A reasonable value for this is best determined by your
# execution environment. When the pool reaches this size, and no idle connections are available,
# calls to getConnection() will block for up to connectionTimeout milliseconds before timing out.
# Timeout value used for socket read operations. If reading from the server takes longer than this value,
# the connection is closed. This can be used as both a brute force global query timeout and a method of
# detecting network problems. The timeout is specified in seconds and a value of 0 means that it is disabled.
# Enabled TLS authentication (if it is enabled, 'tlsClientCert', 'tlsClientKey', and 'tlsCaCert'
# are used and 'password' field is not used).
# Password to an encrypted SSL certificate key file.

# List of network locations at which target database resides. The first entry is always tried in the first
# attempt during creation of database connection. If there are multiple entries specified, then other
# locations are used as fallback method in the order in which they are specified.

# database hostname / IP address
# TCP port on which target database listens to incoming connections

# Repairs the Flyway schema history table before Flyway migration. This will perform the following actions:
# 1. Remove any failed migrations on databases.
# 2. Realign the checksums, descriptions and types of the applied migrations with available migrations.
# Name / unique identifier of local Uniconfig instance.
# Hostname or IP address of a local host that runs Uniconfig.

# UniConfig node identification

By default, UniConfig node names are generated randomly. This behavior can be modified by setting db-persistence.uniconfig-instance.instance-name. The instance name is leveraged, for example, in the clustering of stream subscriptions.


node identification properties
# Name / unique identifier of local Uniconfig instance.
# Hostname or IP address of a local host that runs Uniconfig.

# Kafka and notification properties

This section contains properties related to connections to Kafka brokers, Kafka publisher timeouts, authentication, subscription allocation and rebalancing settings.

Example with essential properties:

notification properties
# Grouped settings that are related to notifications.

# Flag that determines whether notifications are collected.
# How many unassigned subscriptions can be processed.
# How many unassigned subscriptions can be processed within one subscription monitoring interval.
# How often should UniConfig check for unassigned notifications subscriptions (in seconds).
# If response body should be included in notification.
# If calculating the diff result should be included in notification.
# Maximum count of records, after reaching this count the oldest records will be deleted.
# Maximum age of records, all older records will be deleted (in hours).

# Grouped settings that are related to kafka.

# Flag that determines whether audit logs are enabled.
# Flag that determines whether netconf notifications are enabled.
# Flag that determines whether gnmi notifications are enabled.
# Flag that determines whether snmp notifications are enabled.
# Flag that sets the port to listen to notifications.
# The time in which the cache entry will expire (in seconds) if not accessed. This cache is used to map addresses
# from notifications to repository names that is used to parse the specific notification.
# The time in which the cache entry will expire (in seconds) if not accessed. This cache is used to map object
# identifiers to translated yang instance identifiers that are used to map them in the UniConfig schema context.
# Flag that determines whether transaction notifications are enabled.
# Enabled collection and propagation of data-change-events into Kafka.
# Enabled collection and propagation of connection notifications into Kafka.
# Unique identifier of topic that is used for storing netconf notifications.
# Unique identifier of topic that is used for storing GNMI notifications.
# Unique identifier of topic that is used for storing SNMP notifications.
# Unique identifier of topic that is used for storing audit logs.
# Unique identifier of topic that is used for storing transaction notifications.
# Unique identifier of the Kafka topic used for distribution of data-change-events.
# Unique identifier of the Kafka topic used for distribution of connection notifications.
# If only connection notifications for NETCONF stream are enabled.
# If only connection notifications for gNMI stream are enabled.
# The maximum thread pool size in the executor.
# A thread pool executor is needed to send messages to Kafka.
# The maximum capacity of the work queue in the executor.
# Unique identifier of the Kafka topic used for distribution of shell notifications.
# Enabled shell notifications into Kafka.
# List of Address / hostname of the interface on which Kafka broker is listening to incoming connections and
# TCP port on which Kafka broker is listening to incoming connections.
# Specifies the number of messages that the Kafka handler processes as a batch.
# In kafka is set with the parameter 'batch.size'.
# Configuration of how long will the producer wait for the acknowledgement of a request. (in ms)
# If the acknowledgement is not received before the timeout elapses, the producer will resend the
# request or fail the request if retries are exhausted.
# In kafka is set with the parameter 'request.timeout.ms'.
# Configuration of the upper bound on the time to report success or failure after a
# call to send() returns.(in ms)
# This limits the total time that a record will be delayed prior to sending, the time to
# await acknowledgement from the broker (if expected), and the time allowed for retriable send failures.
# In kafka is set with the parameter 'delivery.timeout.ms'.
# Configuration of how long the send() method and the creation of connection for
# reading of metadata methods will block. (in ms).
# In kafka is set with the parameter 'max.block.ms'.
# The producer groups together any records that arrive in between request transmissions into a single batched
# request. Normally this occurs only under load when records arrive faster than they can be sent out.
# However, in some circumstances the client may want to reduce the number of requests even under moderate load.
# This setting accomplishes this by adding a small amount of artificial delay—that is, rather than immediately
# sending out a record, the producer will wait for up to the given delay to allow other records to be sent so
# that the sends can be batched together. This can be thought of as analogous to Nagle’s algorithm in TCP.
# This setting gives the upper bound on the delay for batching: once we get 'batch.size' worth of records for a
# partition it will be sent immediately regardless of this setting, however if we have fewer than this many
# bytes accumulated for this partition we will ‘linger’ for the specified time waiting for more records to show
# up. This setting defaults to 0 (i.e. no delay). Setting linger=5, for example, would have the effect of
# reducing the number of requests sent but would add up to 5ms of latency to records sent in the absence of
# load. In kafka is set with the parameter 'linger.ms'.
# Select compression algorithm for kafka producer.
# Can be one of: none gzip snappy lz4 zstd
# Snappy offers good balance and is good for JSON-based documents and logs.
# The maximum time to block during receiving of next NETCONF notifications in milliseconds.
# The maximum time to block during receiving of next data-change-events in milliseconds.
# If this flag is set to 'true', then embedded Kafka is started during boot process.
# Otherwise, all other settings are effectively ignored.
# Directory to which embedded Kafka is downloaded and extracted.
# URL that is used for download of Kafka, if it hasn't been downloaded yet.
# Directory to which embedded Kafka is downloaded and extracted.
# Clean data from previous run before starting of Kafka (= disabled persistence).
# Number of partitions used for created topic.

# Load-balancer configuration

The following YAML code represents a sample Traefik configuration that can be used in the clustered UniConfig deployment (deployment with a single Traefik node). There is one registered entry-point with the uniconfig identifier on port 8181.

traefik configuration
  insecure: true

    address: ":8181"

  docker: {}

Next, you need to configure UniConfig docker containers with Traefik labels. UniConfig nodes are automatically detected by a Traefik container as uniconfig service providers. There is also a URI prefix (/rests), the name of the "sticky cookie" (uniconfig_server_id) and the server port number (8181) where the UniConfig web server listens to incoming HTTP requests.

configuration of UniConfig docker container
    container_name: "uc_1"
      - traefik.enable=true
      - traefik.http.routers.uniconfig.entrypoints=uniconfig
      - traefik.http.routers.uniconfig.rule=PathPrefix(`/rests`)
      - traefik.http.services.uniconfig.loadBalancer.sticky.cookie.name=uniconfig_server_id
      - traefik.http.services.uniconfig.loadbalancer.server.port=8181

Values for all Traefik labels should be the same on all nodes in the cluster. Scaling of the UniConfig service in the cluster (for example, using Docker Swarm tools) is simple when container settings do not change.

# Clustering of NETCONF subscriptions and notifications

When a device is installed with the stream property set, subscriptions for all provided streams are created in database. These subscriptions are always created with the UniConfig instance id set to null, so they can be acquired by any UniConfig from the cluster.

Each UniConfig instance in a cluster uses its own monitoring system to acquire free subscriptions. The monitoring system uses a specialized transaction to lock subscriptions, which prevents other UniConfig instances from locking the same subscriptions. While locking a subscription, the UniConfig instance writes its id into the subscription table for the currently locked subscription, which indicates that this subscription is already acquired by this UniConfig instance. Other instances of UniConfig will then see that this subscription is not available.

# Optimal subscription count and rebalancing

With multiple UniConfig instances working in a cluster, each instance calculates an optimal range of subscriptions to manage.

# Example: optimal range for 5000 subscriptions with 3 nodes in the cluster and margins set to 0.05 and 0.10 = (1750, 1834)

Based on the optimal range and number of currently opened subscriptions, each UniConfig node (while performing a monitoring system iteration) decides whether it should:

  • Acquire additional subscriptions until the optimal range is reached.
  • Once the optimal range is reached, stay put and not acquire any additional subscriptions.
  • Release some of its subscriptions to trigger rebalancing until the optimal range is reached.

If an instance goes down, all of its subscriptions are immediately released and the optimal ranges of other living nodes are adjusted. Managed network devices, and thus subscriptions, are reopened by the rest of the cluster.

The following example illustrates a timeline for a three-node cluster and how many subscriptions are handled by each node: