3 posts tagged with "Cadence Operations"

At Uber, we previously relied on a dynamic configuration service to manually control the number of partitions for scalable tasklists. This configuration approach introduced several operational challenges:

• Error-prone: Manual updates and deployments were required.
• Unresponsive: Adjustments were typically reactive, often triggered by customer reports or observed backlogs.
• Irreversible: Once increased, the number of partitions was rarely decreased due to the complexity of the two-phase process, especially when anticipating future traffic spikes.

To address these issues, we introduced a new component in the Cadence Matching service: Adaptive Tasklist Scaler. This component dynamically monitors tasklist traffic and adjusts partition counts automatically. Since its rollout, we've seen a significant reduction in incidents and operational overhead caused by misconfigured tasklists.

---

What is a Scalable Tasklist?​

A scalable tasklist is one that supports multiple partitions. Since Cadence’s Matching service is sharded by tasklist, all requests to a specific tasklist are routed to a single Matching host. To avoid bottlenecks and enhance scalability, tasklists can be partitioned so that multiple Matching hosts handle traffic concurrently.

These partitions are transparent to clients. When a request arrives at the Cadence server for a scalable tasklist, the server selects an appropriate partition. More details can be found in this document.

How Is the Number of Partitions Manually Configured?​

The number of partitions for a tasklist is controlled by two dynamic configuration properties:

1. matching.numTasklistReadPartitions: Specifies the number of read partitions.
2. matching.numTasklistWritePartitions: Specifies the number of write partitions.

To prevent misconfiguration, a guardrail is in place to ensure that the number of read partitions is never less than the number of write partitions.

When increasing the number of partitions, both properties are typically updated simultaneously. However, due to the guardrail, the order of updates doesn't matter—read and write partitions can be increased in any sequence.

In contrast, decreasing the number of partitions is more complex and requires a two-phase process:

1. First, reduce the number of write partitions.
2. Then, wait for any backlog in the decommissioned partitions to drain completely.
3. Finally, reduce the number of read partitions.

Because this process is tedious, error-prone, and backlog-sensitive, it is rarely performed in production environments.

---

How Does Adaptive Tasklist Scaler Work?​

The architecture of the adaptive tasklist scaler is shown below:

1. Migrating Configuration to the Database​

The first key change was migrating partition count configuration from dynamic config to the Cadence cluster’s database. This allows the configuration to be updated programmatically.

• The adaptive tasklist scaler runs in the root partition only.
• It reads and updates the partition count.
• Updates propagate to non-root partitions via a push model, and to pollers and producers via a pull model.
• A version number is associated with each config. The version only increments through scaler updates, ensuring monotonicity and consistency across components.

2. Monitoring Tasklist Traffic​

The scaler periodically monitors the write QPS of each tasklist.

• If QPS exceeds an upscale threshold for a sustained period, the number of read and write partitions is increased proportionally.
• If QPS falls below a downscale threshold, only the write partitions are reduced initially. The system then waits for drained partitions to clear before reducing the number of read partitions, ensuring backlog-free downscaling.

---

Enabling Adaptive Tasklist Scaler​

Prerequisites​

To use this feature, upgrade Cadence to v1.3.0 or later.

Also, migrate tasklist partition configurations to the database using this guide.

Configuration​

The scaler is governed by the following dynamic configuration parameters:

• matching.enableAdaptiveScaler: Enables the scaler at the tasklist level.
• matching.partitionUpscaleSustainedDuration: Duration that QPS must stay above threshold before triggering upscale.
• matching.partitionDownscaleSustainedDuration: Duration below threshold required before triggering downscale.
• matching.adaptiveScalerUpdateInterval: Frequency at which the scaler evaluates and updates partition counts.
• matching.partitionUpscaleRPS: QPS threshold per partition that triggers upscale.
• matching.partitionDownscaleFactor: Factor applied to introduce hysteresis, lowering the QPS threshold for downscaling to avoid oscillations.

---

Monitoring and Metrics​

Several metrics have been introduced to help monitor the scaler’s behavior:

QPS and Thresholds​

• estimated_add_task_qps_per_tl: Estimated QPS of task additions per tasklist.
• tasklist_partition_upscale_threshold: Upscale threshold for task additions.
• tasklist_partition_downscale_threshold: Downscale threshold for task additions.

The estimated_add_task_qps_per_tl value should remain between the upscale and downscale thresholds. If not, the scaler may not be functioning properly.

Partition Configurations​

• task_list_partition_config_num_read: Number of current read partitions.
• task_list_partition_config_num_write: Number of current write partitions.
• task_list_partition_config_version: Version of the current partition configuration.

These metrics are emitted by various components: root and non-root partitions, pollers, and producers. Their values should align under normal conditions, except immediately after updates.

---

Status at Uber​

We enabled adaptive tasklist scaler across all Uber clusters in March 2025. Since its deployment:

• Zero incidents have been reported due to misconfigured tasklists.
• Operational workload related to manual scaling has been eliminated.
• Scalability and resilience of Matching service have improved significantly.

At Uber, we want to achieve regional resilience such that losing a zone within a region can be tolerated without requiring a cross-region failover. We also want to make sure that losing a zone only affects a subset of workload, at most, rather than everything. However, in Cadence-based systems, the workload in a region is distributed randomly across all workers in the region at a “task-level granularity”, which means a workflow may be worked on by any worker in the region where the domain is active. To achieve this goal, we introduced Zonal Isolation for Cadence Workflows - a feature designed to pin workflows to the zone they are started in, so that zonal isolation can be achieved at a workflow-level.

What is Zonal Isolation for Cadence Workflows?​

At high-level, Zonal Isolation for Cadence Workflows can be thought in 2 levels:

1. Task-level isolation: All decision tasks and activity tasks of a workflow are only processed by workers from the zone where the workflow was started
2. Infrastructure-level isolation: Within a regional Cadence cluster, workflows are handled by server instances in the same zone where they were started, and the corresponding data is stored in that zone as well.

Infrastructure-level isolation is quite challenging to implement as it requires significant changes to the core design of the Cadence server. Due to the complexity involved, support for this feature is not planned for the foreseeable future.

As a result, the focus remains on achieving task-level zonal isolation outside the Cadence server, which offers a more practical and immediate way to improve system resilience. It provides the capability of ensuring that an unhealthy zone (i.e. bad deployment of workers) only affect a subset of workflows (started from a certain zone) rather than every workflow in a Cadence domain.

Background​

Cadence historically has been using TChannel transport with Thrift encoding for both internal RPC calls and communication with client SDKs. gRPC is becoming a de-facto industry standard with much better adoption and community support. It offers features such as authentication and streaming that are very relevant for Cadence. Moreover, TChannel is being deprecated within Uber itself, pushing an effort for this migration. During the last year we’ve implemented multiple changes in server and SDK that allows users to use gRPC in Cadence, as well as to upgrade their existing Cadence cluster in a backward compatible way. This post tracks the completed work items and our future plans.

Our Approach​

With ~500 services using Cadence at Uber and many more open source customers around the world, we had to think about the gRPC transition in a backwards compatible way. We couldn’t simply flip transport and encoding everywhere. Instead we needed to support both protocols as an intermediate step to ensure a smooth transition for our users.

Cadence was using Thrift/TChannel not just for the API with client SDKs. They were also used for RPC calls between internal Cadence server components and also between different data centers. When starting this migration we had a choice of either starting with public APIs first or all the internal things within the server. We chose the latter one, so that we could gain experience and iterate faster within the server without disruption to the clients. With server side done and listening for both protocols, dynamic config flag was exposed to switch traffic internally. It allowed gradual deployment and provided an option to rollback if needed.