Skip to content

011. Device Request Queue

Feature: #1479
Authors: Atif Aziz
Status: Accepted

Context

LoRa devices, which can reach multiple LNSs, pose an interesting challenge for Edge Hub since Edge Hub wasn't designed for devices competing for gateways. When a device connection is made to IoT Hub, Edge Hub continues to maintain that connection in the open state (for network resiliency) until explicitly closed. Moreover, IoT Hub allows maintining only a single connection per device. When another is opened, the existing one is closed. If a device can reach multiple LNSs then Edge Hub and IoT Hub begin an aggressive ping-pong of connections being opened and closed. In other words, as each LNS or Edge Hub tries to open a connection for the same device, IoT Hub closes the connection for the other. But then the other tries to re-open the connection and the ping-pong continues. This problem needs addressing to avoid scalability limitations.

The ping-pong can be prevented by electing a leader LNS for a device through the function race determination and as a fallback announcing the decision to others using a C2D message. An LNS receiving such a message will then close that connection so that Edge Hub doesn't continue to maintain the connection in the open state through retires. The handling of the message needs to know:

  • It can close the connection right now, without interfering with a currently running operation, such as a message delivery where the message could be potentially dropped.
  • There are no operations starting around the same time that would cause it to be opened again.

The following cases can trigger operations from different threads:

  • An OTAA join message; DeviceJoinLoader eventually loads the twins and stores them back up.
  • A data message; DeviceLoaderSynchronizer loads the twins for multiple devices based on the DevAddr and adds them to the cache.
  • C2D message for a class C device that needs to be delivered can trigger a load if the device is not in the cache.
  • Cache refreshes can happen on a background thread for a device, causing the need for a connection.

There are a number of complex and concurrent code paths in the current solution that make it hard to reason about when a device connection is in use, when it's safe to open/close without affecting another operation in flight and how various device-related operations (such as refreshing from the twin) can work in isolation, deterministically and without race conditions. If a single queue could be maintained for all device-related operations then it would become easier to order those operations, reason about them and make they don't cause connections to be opened when an LNS has lost the race against another.

Several approaches were explored to understand the overall impact of refactoring, whether the changes to the current code base and design would be large or small and worth the benefits they bring. The approaches explored can be summed up as follows:

  1. Add a queue to LoRaDevice for all device-related operations. There is one that exists today but it is used to service uplink data requests only. The idea would be to extend it to encompass all other requests and operations.

  2. Turn the LNS implementation to be (logically) single-threaded so that not only does it make it easier to reason about the code (except perhaps with regards to re-entrancy), but it also enables use of very simple data structures without the need for locking. This is possible because the bulk of LNS is I/O bound. There is practically no CPU-intensive code that executes between each await.

  3. Simplify the entire flow and processing of a message into a simple request-response model that can be manipulated and reasoned about much more easily. That is, all methods receiving and processing a LoRaRequest return Task<DownlinkMessage> instead of void or Task. This allows all Task.Run uses to be moved to a single and central point when a message is received instead of being littered throughout the code base. The main message loop could also be made responsible for central error-handling/logging and sending of downlink messages when the processing of a request has completed. This would just require regular use of tasks. Next, all Task-based operations on LoRaDevice could be naturally and implicitly queued and then executed in a mutually exclusive manner.

Decision

All operations for a particular device requiring the IoT Hub connection are to be executed in a serial/exclusive fashion on the LoRaDevice itself. This will ensure that there is only one operation acting on the connection at any given time. Making each execution exclusive (think queue) gives a deterministic way of adding a close operation without affecting any other operations that may be in flight.

Unfortunately, the initialization of a LoRaDevice cannot be addressed with the same approach since it requires a two-step initialization per its current design. That is, a LoRaDevice is minimally initialized when created so its InitializeAsync method can be called to fetch the twin information. Until that completes, a LoRaDevice is not considered technically ready for being put in the cache. Since a LoRaDevice cannot be discovered during its initialization, it cannot be synchronized with the other operations. To work around this, load operations will be synchronized in the LoRaDeviceRegistry to ensure there is only ever one operation loading the twins for a particular DevEUI. It was also decided that closing the connection after initialization is the easiest approach to delay the connection ownership decision to after the arrival of the first message.

The approach to simplify the entire flow and processing of a message into a simple request-response model will be taken as a stretch goal so that the connection management can be implemented and tested earlier. A quick spike demonstrated that the changes to the main code base would be fairly contained and the largest impact is expected to be in adapting the tests (which could also be done with a stop-gap measure where the tests are adapted after the initial refactoring of the code base).

Based on a spike of the changes the simple request-response model would require, it seems plausible to achieve the refactoring within (at most) two weekly sprints. It is expected that the bulk of the time will be spent in refactoring the test code.

Consequences

The chosen approach has the following benefits:

  • Single use of Task.Run in LnsProtocolMessageProcessor. This reduces the number execution forks to consider as well as chances of exceptions going unobserved.
  • Removes many abstractions like ILoRaDeviceRequestQueue and IMessageDispatcher, implementations like DeviceLoaderSynchronizer and ExternalGatewayLoRaRequestQueue, and potentially more.
  • Easier to reason about the overall message flow.
  • Use of the regular async-await programming model, including error-handling and cancellation that's built-in into tasks.
  • Transparent use of queues through tasks in LoRaDevice to serialize execution. This subsumes the first approach (discussed in the introductory section) by having implicit rather explicit queuing (at the application-level).
  • Greatest potential to simplify tests since assertions can rely on simply return values and exceptions, instead of success/failure notifications.
  • It is not mutually exclusive with other approaches explored. For example, by simplifying to the request-response model, it would be even easier to have the LNS operate with a single logical thread if that could further help remove some complexity (without compromising scalability) like locks.

The absence of an explicit queue could be make it more difficult for someone to understand the code and choices made if they are not familiar with the intricacies of how async-await operates.


Last update: 2022-02-24
Created: 2022-02-23