Migrating Crucial Site visitors At Scale with No Downtime — Half 2 | by Netflix Expertise Weblog | Might, 2023

Shyam Gala, Javier Fernandez-Ivern, Anup Rokkam Pratap, Devang Shah

Image your self enthralled by the most recent episode of the one that you love Netflix sequence, delighting in an uninterrupted, high-definition streaming expertise. Behind these good moments of leisure is a posh mechanism, with quite a few gears and cogs working in concord. However what occurs when this equipment wants a change? That is the place large-scale system migrations come into play. Our earlier weblog publish offered replay site visitors testing — a vital instrument in our toolkit that enables us to implement these transformations with precision and reliability.

Replay site visitors testing provides us the preliminary basis of validation, however as our migration course of unfolds, we’re met with the necessity for a fastidiously managed migration course of. A course of that doesn’t simply reduce threat, but additionally facilitates a steady analysis of the rollout’s impression. This weblog publish will delve into the methods leveraged at Netflix to introduce these modifications to manufacturing.

Canary deployments are an efficient mechanism for validating modifications to a manufacturing backend service in a managed and restricted method, thus mitigating the danger of unexpected penalties which will come up because of the change. This course of entails creating two new clusters for the up to date service; a baseline cluster containing the present model operating in manufacturing and a canary cluster containing the brand new model of the service. A small share of manufacturing site visitors is redirected to the 2 new clusters, permitting us to watch the brand new model’s efficiency and examine it in opposition to the present model. By amassing and analyzing key efficiency metrics of the service over time, we are able to assess the impression of the brand new modifications and decide in the event that they meet the provision, latency, and efficiency necessities.

Some product options require a lifecycle of requests between the client machine and a set of backend companies to drive the characteristic. As an example, video playback performance on Netflix entails requesting URLs for the streams from a service, calling the CDN to obtain the bits from the streams, requesting a license to decrypt the streams from a separate service, and sending telemetry indicating the profitable begin of playback to yet one more service. By monitoring metrics solely on the stage of service being up to date, we’d miss capturing deviations in broader end-to-end system performance.

Sticky Canary is an enchancment to the standard canary course of that addresses this limitation. On this variation, the canary framework creates a pool of distinctive buyer units after which routes site visitors for this pool persistently to the canary and baseline clusters all through the experiment. Aside from measuring service-level metrics, the canary framework is ready to maintain observe of broader system operational and buyer metrics throughout the canary pool and thereby detect regressions on all the request lifecycle circulate.

Sticky Canary

It is very important notice that with sticky canaries, units within the canary pool proceed to be routed to the canary all through the experiment, doubtlessly leading to undesirable habits persisting by way of retries on buyer units. Due to this fact, the canary framework is designed to watch operational and buyer KPI metrics to detect persistent deviations and terminate the canary experiment if obligatory.

Canaries and sticky canaries are useful instruments within the system migration course of. In comparison with replay testing, canaries enable us to increase the validation scope past the service stage. They permit verification of the broader end-to-end system performance throughout the request lifecycle for that performance, giving us confidence that the migration is not going to trigger any disruptions to the client expertise. Canaries additionally present a possibility to measure system efficiency below totally different load circumstances, permitting us to establish and resolve any efficiency bottlenecks. They permit us to additional fine-tune and configure the system, guaranteeing the brand new modifications are built-in easily and seamlessly.

A/B testing is a well known methodology for verifying hypotheses by way of a managed experiment. It entails dividing a portion of the inhabitants into two or extra teams, every receiving a special therapy. The outcomes are then evaluated utilizing particular metrics to find out whether or not the speculation is legitimate. The trade steadily employs the method to evaluate hypotheses associated to product evolution and person interplay. Additionally it is extensively utilized at Netflix to check modifications to product habits and buyer expertise.

A/B testing can be a useful software for assessing vital modifications to backend techniques. We will decide A/B take a look at membership in both machine software or backend code and selectively invoke new code paths and companies. Inside the context of migrations, A/B testing permits us to restrict publicity to the migrated system by enabling the brand new path for a smaller share of the member base. Thereby controlling the danger of surprising habits ensuing from the brand new modifications. A/B testing can be a key method in migrations the place the updates to the structure contain altering machine contracts as properly.

Canary experiments are usually performed over intervals starting from hours to days. Nevertheless, in sure situations, migration-related experiments could also be required to span weeks or months to acquire a extra correct understanding of the impression on particular High quality of Expertise (QoE) metrics. Moreover, in-depth analyses of explicit enterprise Key Efficiency Indicators (KPIs) might require longer experiments. As an example, envision a migration situation the place we improve the playback high quality, anticipating that this enchancment will result in extra clients partaking with the play button. Assessing related metrics throughout a substantial pattern dimension is essential for acquiring a dependable and assured analysis of the speculation. A/B frameworks work as efficient instruments to accommodate this subsequent step within the confidence-building course of.

Along with supporting prolonged durations, A/B testing frameworks provide different supplementary capabilities. This method permits take a look at allocation restrictions based mostly on elements similar to geography, machine platforms, and machine variations, whereas additionally permitting for evaluation of migration metrics throughout related dimensions. This ensures that the modifications don’t disproportionately impression particular buyer segments. A/B testing additionally gives adaptability, allowing changes to allocation dimension all through the experiment.

We would not use A/B testing for each backend migration. As a substitute, we use it for migrations wherein modifications are anticipated to impression machine QoE or enterprise KPIs considerably. For instance, as mentioned earlier, if the deliberate modifications are anticipated to enhance consumer QoE metrics, we’d take a look at the speculation by way of A/B testing.

After finishing the varied phases of validation, similar to replay testing, sticky canaries, and A/B exams, we are able to confidently assert that the deliberate modifications is not going to considerably impression SLAs (service-level-agreement), machine stage QoE, or enterprise KPIs. Nevertheless, it’s crucial that the ultimate rollout is regulated to make sure that any unnoticed and surprising issues don’t disrupt the client expertise. To this finish, we have now applied site visitors dialing because the final step in mitigating the danger related to enabling the modifications in manufacturing.

A dial is a software program assemble that permits the managed circulate of site visitors inside a system. This assemble samples inbound requests utilizing a distribution perform and determines whether or not they need to be routed to the brand new path or stored on the prevailing path. The choice-making course of entails assessing whether or not the distribution perform’s output aligns inside the vary of the predefined goal share. The sampling is completed persistently utilizing a hard and fast parameter related to the request. The goal share is managed by way of a globally scoped dynamic property that may be up to date in real-time. By growing or lowering the goal share, site visitors circulate to the brand new path will be regulated instantaneously.


The choice of the particular sampling parameter relies on the particular migration necessities. A dial can be utilized to randomly pattern all requests, which is achieved by deciding on a variable parameter like a timestamp or a random quantity. Alternatively, in situations the place the system path should stay fixed with respect to buyer units, a continuing machine attribute similar to deviceId is chosen because the sampling parameter. Dials will be utilized in a number of locations, similar to machine software code, the related server part, and even on the API gateway for edge API techniques, making them a flexible software for managing migrations in advanced techniques.

Site visitors is dialed over to the brand new system in measured discrete steps. At each step, related stakeholders are knowledgeable, and key metrics are monitored, together with service, machine, operational, and enterprise metrics. If we uncover an surprising challenge or discover metrics trending in an undesired path throughout the migration, the dial provides us the potential to shortly roll again the site visitors to the outdated path and tackle the problem.

The dialing steps will also be scoped on the knowledge middle stage if site visitors is served from a number of knowledge facilities. We will begin by dialing site visitors in a single knowledge middle to permit for a better side-by-side comparability of key metrics throughout knowledge facilities, thereby making it simpler to watch any deviations within the metrics. The length of how lengthy we run the precise discrete dialing steps will also be adjusted. Operating the dialing steps for longer intervals will increase the chance of surfacing points which will solely have an effect on a small group of members or units and might need been too low to seize and carry out shadow site visitors evaluation. We will full the ultimate step of migrating all of the manufacturing site visitors to the brand new system utilizing the mix of gradual step-wise dialing and monitoring.

Stateful APIs pose distinctive challenges that require totally different methods. Whereas the replay testing method mentioned within the earlier a part of this weblog sequence will be employed, extra measures outlined earlier are obligatory.

This alternate migration technique has confirmed efficient for our techniques that meet sure standards. Particularly, our knowledge mannequin is straightforward, self-contained, and immutable, with no relational features. Our system doesn’t require strict consistency ensures and doesn’t use database transactions. We undertake an ETL-based dual-write technique that roughly follows this sequence of steps:

  • Preliminary Load by way of an ETL course of: Knowledge is extracted from the supply knowledge retailer, remodeled into the brand new mannequin, and written to the newer knowledge retailer by way of an offline job. We use customized queries to confirm the completeness of the migrated information.
  • Steady migration by way of Twin-writes: We make the most of an active-active/dual-writes technique to migrate the majority of the information. As a security mechanism, we use dials (mentioned beforehand) to manage the proportion of writes that go to the brand new knowledge retailer. To keep up state parity throughout each shops, we write all state-altering requests of an entity to each shops. That is achieved by deciding on a sampling parameter that makes the dial sticky to the entity’s lifecycle. We incrementally flip the dial up as we achieve confidence within the system whereas fastidiously monitoring its general well being. The dial additionally acts as a change to show off all writes to the brand new knowledge retailer if obligatory.
  • Steady verification of information: When a document is learn, the service reads from each knowledge shops and verifies the practical correctness of the brand new document if present in each shops. One can carry out this comparability dwell on the request path or offline based mostly on the latency necessities of the actual use case. Within the case of a dwell comparability, we are able to return information from the brand new datastore when the information match. This course of provides us an thought of the practical correctness of the migration.
  • Analysis of migration completeness: To confirm the completeness of the information, chilly storage companies are used to take periodic knowledge dumps from the 2 knowledge shops and in contrast for completeness. Gaps within the knowledge are crammed again with an ETL course of.
  • Minimize-over and clean-up: As soon as the information is verified for correctness and completeness, twin writes and reads are disabled, any consumer code is cleaned up, and skim/writes solely happen to the brand new knowledge retailer.
Migrating Stateful Methods

Clear-up of any migration-related code and configuration after the migration is essential to make sure the system runs easily and effectively and we don’t construct up tech debt and complexity. As soon as the migration is full and validated, all migration-related code, similar to site visitors dials, A/B exams, and replay site visitors integrations, will be safely faraway from the system. This contains cleansing up configuration modifications, reverting to the unique settings, and disabling any momentary elements added throughout the migration. As well as, it is very important doc all the migration course of and maintain information of any points encountered and their decision. By performing a radical clean-up and documentation course of, future migrations will be executed extra effectively and successfully, constructing on the teachings discovered from the earlier migrations.

We now have utilized a spread of methods outlined in our weblog posts to conduct quite a few giant, medium, and small-scale migrations on the Netflix platform. Our efforts have been largely profitable, with minimal to no downtime or vital points encountered. All through the method, we have now gained useful insights and refined our methods. It must be famous that not all the methods offered are universally relevant, as every migration presents its personal distinctive set of circumstances. Figuring out the suitable stage of validation, testing, and threat mitigation requires cautious consideration of a number of elements, together with the character of the modifications, potential impacts on buyer expertise, engineering effort, and product priorities. Finally, we purpose to attain seamless migrations with out disruptions or downtime.

In a sequence of forthcoming weblog posts, we’ll discover a collection of particular use circumstances the place the methods highlighted on this weblog sequence have been utilized successfully. They’ll deal with a complete evaluation of the Adverts Tier Launch and an in depth GraphQL migration for numerous product APIs. These posts will provide readers invaluable insights into the sensible software of those methodologies in real-world conditions.