RBFS API Fundamentals

RtBrick Full Stack (RBFS) architecture basics and an understanding of the RBFS API characteristics is needed to make use of the RBFS APIs. The Brick Data Store (BDS) is a schema-driven in-memory database optimized for networking use cases. It stores objects, which are described in schemas, and organizes them in different tables. Each brick daemon has a single responsibility and defines the tables and schemas needed for its respective task. Brick daemons interact with each other over a pub-sub facility provided by the BDS.

RBFS differentiates between persistent and non-persistent configuration. Non-persistent configuration is lost after each switch-reboot whereas persistent configuration is written to disk. Service related information is an example for non-persistent configuration. All service information is lost when a switch is rebooted and has to be pushed again on demand by the controller after the reboot. The startup configuration is an example for persistent data. Whenever startup related configuration gets changed, e.g. by adding a new BGP peering, the startup configuration file gets updated to preserve the new settings if the switch reboots.

The BDS API provides access to the BDS tables and objects. The BDS API relies on RPC over HTTP and defines an envelope protocol to convey BDS objects. The data exchange format is JSON.

The BDS API supports adding objects, removing objects and walking through a table. Working with BDS tables and objects requires awareness of the indices that exist on each table. The add operation adds new or replaces existing objects.

The delete operation removes objects. A merge operation to add and remove objects in one go does not exist. Walking a table requires to specify the index the walk operation should follow. The sequence index is used by default if no index is specified. The sequence index represents the temporal insertion order of objects into a table. The objects are returned in the index sort order. Searching for objects requires an understanding of key and non-key attributes of each index.

The usage of the BDS API also requires an understanding of the relations between the objects. It can be compared to referential integrity in SQL databases which also forces the client to modify database records in the right order. Using the BDS API for integration is as if you would use SQL (or the database model) rather than an API of a software product relying on a relational database.

For example, BDS schemas definitions can be modified to simplify CLI implementation without changing the data that is being exchanged as such. Even a patch can introduce a non-backward compatible BDS schema definition change.

For every brick daemon there is a dedicated BDS API endpoint.

Every brick daemon provides additional APIs beside the BDS API. The configuration daemon is a good example. It introduces a transactional configuration API that allows you to incrementally update a candidate configuration that is later applied in one go. The configuration daemon detects deltas between configurations and only the detected deltas trigger subsequent actions. For example, if a new BGP peer has been added to the candidate config, the additional BGP peering gets established without resetting established peerings.

All configurations applied via config daemon will be added to the startup configuration file and applied after every reboot. Not all BDS tables become part of the persistent configuration by default. A brick daemon can declare a BDS table persistent. In this case the ownership of the table is transferred from the respective brick daemon to the configuration daemon. The brick daemon cannot modify the objects in BDS tables when the ownership has been transferred to configuration daemon. Consequently, the BDS low-level API can still be used for reading objects but not for object modifications. All changes are applied by the configuration daemon. The configuration daemon provides a delta processing and the brick daemons get notified about configuration changes via the BDS pub-sub facility mentioned earlier.

The YANG models describe the configuration syntax. The RESTCONF API provides the configuration data only, while the state information is provided by the Operational State API.

The Operational State API forms a stable contract between network management systems and RBFS. The API is backward compatible to support running multiple versions of RBFS in the network, which is typically the case when a new RBFS release is rolled out in the network. In addition, the formal Operational State API contract simplifies developing management system integrations by using API mock-ups rather than running RBFS instances with the respective services configured. This lowers the entry-barrier for network management integration for both, customers and integration partners. The Operational State API is implemented in Python which allows customers, integration partners, professional services and engineering to collaborate on the API endpoints.

The control daemon (CTRLD) is the entry point to RBFS. In addition, CTRLD implements the ZTP startup configuration discovery by means of downloading the startup configuration files from the retrieved base URL and applying it to the switch.

CTRLD exposes a REST-API to program the switch, which is kept backward compatible /CTRLD/. CTRLD itself uses the low-level BDS API as needed. For example, CTRLD API supports to specify which metric shall be sampled and to define alert conditions which are then stored in BDS tables. CTRLD takes care of compatibility issues outlined earlier. RBFS leverages Prometheus for metric sampling. All counters and metrics of interest shall be sampled by Prometheus and accessed through PromQL rather than the BDS API. A local history of all sampled metrics is also very handy for troubleshooting.

The CTRLD API also provides access to the low-level BDS API endpoints. The usage of the BDS API is an indication that a CTRLD API endpoint is missing. It is recommended to define an external representation of state implemented by the CTRLD API rather than extensively working with the BDS objects in order to avoid compatibility issues. Since CTRLD runs on the switch it needs only to support the current BDS schema definition to implicitly provide backward compatibility.

The subscriber session establishment relies on RADIUS. All per-subscriber settings are exchanged through RADIUS messages whereas information that is shared across multiple subscribers, like QoS profiles, need to be configured through API invocations.

Subscriber session settings are non-persistent and are lost if a switch reboots or the session ends. The sessions must be established again after each reboot. RADIUS relies on UDP for maximum throughput and minimum latency. Message loss is handled by the application rather than the transport layer. The application asks for the message again if the message is not delivered in time. In addition the application can cope with receiving the same message multiple times.

The previous sections provided a quick overview of the existing RBFS APIs and what they are intended for: