May 18, 2024

by Binbing Hou, Stephanie Vezich Tamayo, Xiao Chen, Liang Tian, Troy Ristow, Haoyuan Wang, Snehal Chennuru, Pawan Dixit

That is the primary of the sequence of our work at Netflix on leveraging knowledge insights and Machine Studying (ML) to enhance the operational automation across the efficiency and value effectivity of massive knowledge jobs. Operational automation–together with however not restricted to, auto analysis, auto remediation, auto configuration, auto tuning, auto scaling, auto debugging, and auto testing–is essential to the success of contemporary knowledge platforms. On this weblog submit, we current our challenge on Auto Remediation, which integrates the at the moment used rule-based classifier with an ML service and goals to mechanically remediate failed jobs with out human intervention. We’ve deployed Auto Remediation in manufacturing for dealing with reminiscence configuration errors and unclassified errors of Spark jobs and noticed its effectivity and effectiveness (e.g., mechanically remediating 56% of reminiscence configuration errors and saving 50% of the financial prices brought on by all errors) and nice potential for additional enhancements.

At Netflix, lots of of 1000’s of workflows and thousands and thousands of jobs are working per day throughout a number of layers of the massive knowledge platform. Given the in depth scope and complicated complexity inherent to such a distributed, large-scale system, even when the failed jobs account for a tiny portion of the full workload, diagnosing and remediating job failures could cause appreciable operational burdens.

For environment friendly error dealing with, Netflix developed an error classification service, referred to as Pensive, which leverages a rule-based classifier for error classification. The rule-based classifier classifies job errors based mostly on a set of predefined guidelines and supplies insights for schedulers to resolve whether or not to retry the job and for engineers to diagnose and remediate the job failure.

Nevertheless, because the system has elevated in scale and complexity, the rule-based classifier has been dealing with challenges attributable to its restricted assist for operational automation, particularly for dealing with reminiscence configuration errors and unclassified errors. Due to this fact, the operational price will increase linearly with the variety of failed jobs. In some instances–for instance, diagnosing and remediating job failures brought on by Out-Of-Reminiscence (OOM) errors–joint effort throughout groups is required, involving not solely the customers themselves, but additionally the assist engineers and area consultants.

To handle these challenges, we’ve developed a brand new function, referred to as Auto Remediation, which integrates the rule-based classifier with an ML service. Based mostly on the classification from the rule-based classifier, it makes use of an ML service to foretell retry success likelihood and retry price and selects the very best candidate configuration as suggestions; and a configuration service to mechanically apply the suggestions. Its main benefits are beneath:

  • Built-in intelligence. As a substitute of fully deprecating the present rule-based classifier, Auto Remediation integrates the classifier with an ML service in order that it may well leverage the deserves of each: the rule-based classifier supplies static, deterministic classification outcomes per error class, which relies on the context of area consultants; the ML service supplies performance- and cost-aware suggestions per job, which leverages the facility of ML. With the built-in intelligence, we are able to correctly meet the necessities of remediating totally different errors.
  • Totally automated. The pipeline of classifying errors, getting suggestions, and making use of suggestions is totally automated. It supplies the suggestions along with the retry resolution to the scheduler, and significantly makes use of a web-based configuration service to retailer and apply really useful configurations. On this means, no human intervention is required within the remediation course of.
  • Multi-objective optimizations. Auto Remediation generates suggestions by contemplating each efficiency (i.e., the retry success likelihood) and compute price effectivity (i.e., the financial prices of working the job) to keep away from blindly recommending configurations with extreme useful resource consumption. For instance, for reminiscence configuration errors, it searches a number of parameters associated to the reminiscence utilization of job execution and recommends the mix that minimizes a linear mixture of failure likelihood and compute price.

These benefits have been verified by the manufacturing deployment for remediating Spark jobs’ failures. Our observations point out that Auto Remediation can efficiently remediate about 56% of all reminiscence configuration errors by making use of the really useful reminiscence configurations on-line with out human intervention; and in the meantime scale back the price of about 50% attributable to its skill to suggest new configurations to make reminiscence configurations profitable and disable pointless retries for unclassified errors. We’ve additionally famous an amazing potential for additional enchancment by mannequin tuning (see the part of Rollout in Manufacturing).


Determine 1 illustrates the error classification service, i.e., Pensive, within the knowledge platform. It leverages the rule-based classifier and consists of three elements:

  • Log Collector is chargeable for pulling logs from totally different platform layers for error classification (e.g., the scheduler, job orchestrator, and compute clusters).
  • Rule Execution Engine is chargeable for matching the collected logs in opposition to a set of predefined guidelines. A rule contains (1) the identify, supply, log, and abstract, of the error and whether or not the error is restartable; and (2) the regex to establish the error from the log. For instance, the rule with the identify SparkDriverOOM contains the data indicating that if the stdout log of a Spark job can match the regex SparkOutOfMemoryError:, then this error is assessed to be a person error, not restartable.
  • End result Finalizer is chargeable for finalizing the error classification outcome based mostly on the matched guidelines. If one or a number of guidelines are matched, then the classification of the primary matched rule determines the ultimate classification outcome (the rule precedence is decided by the rule ordering, and the primary rule has the very best precedence). Alternatively, if no guidelines are matched, then this error will probably be thought of unclassified.


Whereas the rule-based classifier is easy and has been efficient, it’s dealing with challenges attributable to its restricted skill to deal with the errors brought on by misconfigurations and classify new errors:

  • Reminiscence configuration errors. The principles-based classifier supplies error classification outcomes indicating whether or not to restart the job; nevertheless, for non-transient errors, it nonetheless depends on engineers to manually remediate the job. Essentially the most notable instance is reminiscence configuration errors. Such errors are typically brought on by the misconfiguration of job reminiscence. Setting an excessively small reminiscence may end up in Out-Of-Reminiscence (OOM) errors whereas setting an excessively massive reminiscence can waste cluster reminiscence sources. What’s tougher is that some reminiscence configuration errors require altering the configurations of a number of parameters. Thus, setting a correct reminiscence configuration requires not solely the handbook operation but additionally the experience of Spark job execution. As well as, even when a job’s reminiscence configuration is initially properly tuned, modifications akin to knowledge dimension and job definition could cause efficiency to degrade. Provided that about 600 reminiscence configuration errors per thirty days are noticed within the knowledge platform, well timed remediation of reminiscence configuration errors alone requires non-trivial engineering efforts.
  • Unclassified errors. The rule-based classifier depends on knowledge platform engineers to manually add guidelines for recognizing errors based mostly on the recognized context; in any other case, the errors will probably be unclassified. Because of the migrations of various layers of the information platform and the variety of purposes, present guidelines will be invalid, and including new guidelines requires engineering efforts and likewise is dependent upon the deployment cycle. Greater than 300 guidelines have been added to the classifier, but about 50% of all failures stay unclassified. For unclassified errors, the job could also be retried a number of instances with the default retry coverage. If the error is non-transient, these failed retries incur pointless job working prices.


To handle the above-mentioned challenges, our primary methodology is to combine the rule-based classifier with an ML service to generate suggestions, and use a configuration service to use the suggestions mechanically:

  • Producing suggestions. We use the rule-based classifier as the primary go to categorise all errors based mostly on predefined guidelines, and the ML service because the second go to offer suggestions for reminiscence configuration errors and unclassified errors.
  • Making use of suggestions. We use a web-based configuration service to retailer and apply the really useful configurations. The pipeline is totally automated, and the companies used to generate and apply suggestions are decoupled.

Service Integrations

Determine 2 illustrates the mixing of the companies producing and making use of the suggestions within the knowledge platform. The main companies are as follows:

  • Nightingale is a service working the ML mannequin skilled utilizing Metaflow and is chargeable for producing a retry suggestion. The advice contains (1) whether or not the error is restartable; and (2) in that case, the really useful configurations to restart the job.
  • ConfigService is a web-based configuration service. The really useful configurations are saved in ConfigService as a JSON patch with a scope outlined to specify the roles that may use the really useful configurations. When Scheduler calls ConfigService to get really useful configurations, Scheduler passes the unique configurations to ConfigService and ConfigService returns the mutated configurations by making use of the JSON patch to the unique configurations. Scheduler can then restart the job with the mutated configurations (together with the really useful configurations).
  • Pensive is an error classification service that leverages the rule-based classifier. It calls Nightingale to get suggestions and shops the suggestions to ConfigService in order that it may be picked up by Scheduler to restart the job.
  • Scheduler is the service scheduling jobs (our present implementation is with Netflix Maestro). Every time when a job fails, it calls Pensive to get the error classification to resolve whether or not to restart a job and calls ConfigServices to get the really useful configurations for restarting the job.

Determine 3 illustrates the sequence of service calls with Auto Remediation:

  1. Upon a job failure, Scheduler calls Pensive to get the error classification.
  2. Pensive classifies the error based mostly on the rule-based classifier. If the error is recognized to be a reminiscence configuration error or an unclassified error, it calls Nightingale to get suggestions.
  3. With the obtained suggestions, Pensive updates the error classification outcome and saves the really useful configurations to ConfigService; after which returns the error classification outcome to Scheduler.
  4. Based mostly on the error classification outcome acquired from Pensive, Scheduler determines whether or not to restart the job.
  5. Earlier than restarting the job, Scheduler calls ConfigService to get the really useful configuration and retries the job with the brand new configuration.


The ML service, i.e., Nightingale, goals to generate a retry coverage for a failed job that trades off between retry success likelihood and job working prices. It consists of two main elements:

  • A prediction mannequin that collectively estimates a) likelihood of retry success, and b) retry price in {dollars}, conditional on properties of the retry.
  • An optimizer which explores the Spark configuration parameter house to suggest a configuration which minimizes a linear mixture of retry failure likelihood and value.

The prediction mannequin is retrained offline each day, and is known as by the optimizer to judge every candidate set of configuration parameter values. The optimizer runs in a RESTful service which is known as upon job failure. If there’s a possible configuration resolution from the optimization, the response contains this suggestion, which ConfigService makes use of to mutate the configuration for the retry. If there is no such thing as a possible resolution–in different phrases, it’s unlikely the retry will succeed by altering Spark configuration parameters alone–the response features a flag to disable retries and thus get rid of wasted compute price.

Prediction Mannequin

Provided that we need to discover how retry success and retry price may change underneath totally different configuration eventualities, we want some solution to predict these two values utilizing the data we’ve concerning the job. Information Platform logs each retry success consequence and execution price, giving us dependable labels to work with. Since we use a shared function set to foretell each targets, have good labels, and have to run inference shortly on-line to satisfy SLOs, we determined to formulate the issue as a multi-output supervised studying activity. Specifically, we use a easy Feedforward Multilayer Perceptron (MLP) with two heads, one to foretell every consequence.

Coaching: Every report within the coaching set represents a possible retry which beforehand failed attributable to reminiscence configuration errors or unclassified errors. The labels are: a) did retry fail, b) retry price. The uncooked function inputs are largely unstructured metadata concerning the job such because the Spark execution plan, the person who ran it, and the Spark configuration parameters and different job properties. We break up these options into these that may be parsed into numeric values (e.g., Spark executor reminiscence parameter) and those who can’t (e.g., person identify). We used function hashing to course of the non-numeric values as a result of they arrive from a excessive cardinality and dynamic set of values. We then create a decrease dimensionality embedding which is concatenated with the normalized numeric values and handed by means of a number of extra layers.

Inference: Upon passing validation audits, every new mannequin model is saved in Metaflow Internet hosting, a service supplied by our inside ML Platform. The optimizer makes a number of calls to the mannequin prediction perform for every incoming configuration suggestion request, described in additional element beneath.


When a job try fails, it sends a request to Nightingale with a job identifier. From this identifier, the service constructs the function vector for use in inference calls. As described beforehand, a few of these options are Spark configuration parameters that are candidates to be mutated (e.g., spark.executor.reminiscence, spark.executor.cores). The set of Spark configuration parameters was based mostly on distilled information of area consultants who work on Spark efficiency tuning extensively. We use Bayesian Optimization (applied by way of Meta’s Ax library) to discover the configuration house and generate a suggestion. At every iteration, the optimizer generates a candidate parameter worth mixture (e.g., spark.executor.reminiscence=7192 mb, spark.executor.cores=8), then evaluates that candidate by calling the prediction mannequin to estimate retry failure likelihood and value utilizing the candidate configuration (i.e., mutating their values within the function vector). After a set variety of iterations is exhausted, the optimizer returns the “greatest” configuration resolution (i.e., that which minimized the mixed retry failure and value goal) for ConfigService to make use of whether it is possible. If no possible resolution is discovered, we disable retries.

One draw back of the iterative design of the optimizer is that any bottleneck can block completion and trigger a timeout, which we initially noticed in a non-trivial variety of instances. Upon additional profiling, we discovered that many of the latency got here from the candidate generated step (i.e., determining which instructions to step within the configuration house after the earlier iteration’s analysis outcomes). We discovered that this situation had been raised to Ax library house owners, who added GPU acceleration options in their API. Leveraging this selection decreased our timeout charge considerably.

We’ve deployed Auto Remediation in manufacturing to deal with reminiscence configuration errors and unclassified errors for Spark jobs. In addition to the retry success likelihood and value effectivity, the affect on person expertise is the main concern:

  • For reminiscence configuration errors: Auto remediation improves person expertise as a result of the job retry isn’t profitable and not using a new configuration for reminiscence configuration errors. Which means that a profitable retry with the really useful configurations can scale back the operational masses and save job working prices, whereas a failed retry doesn’t make the person expertise worse.
  • For unclassified errors: Auto remediation recommends whether or not to restart the job if the error can’t be labeled by present guidelines within the rule-based classifier. Specifically, if the ML mannequin predicts that the retry may be very prone to fail, it should suggest disabling the retry, which might save the job working prices for pointless retries. For instances through which the job is business-critical and the person prefers all the time retrying the job even when the retry success likelihood is low, we are able to add a brand new rule to the rule-based classifier in order that the identical error will probably be labeled by the rule-based classifier subsequent time, skipping the suggestions of the ML service. This presents some great benefits of the built-in intelligence of the rule-based classifier and the ML service.

The deployment in manufacturing has demonstrated that Auto Remediation can present efficient configurations for reminiscence configuration errors, efficiently remediating about 56% of all reminiscence configuration with out human intervention. It additionally decreases compute price of those jobs by about 50% as a result of it may well both suggest new configurations to make the retry profitable or disable pointless retries. As tradeoffs between efficiency and value effectivity are tunable, we are able to resolve to realize the next success charge or extra price financial savings by tuning the ML service.

It’s value noting that the ML service is at the moment adopting a conservative coverage to disable retries. As mentioned above, that is to keep away from the affect on the instances that customers desire all the time retrying the job upon job failures. Though these instances are anticipated and will be addressed by including new guidelines to the rule-based classifier, we contemplate tuning the target perform in an incremental method to steadily disable extra retries is useful to offer fascinating person expertise. Given the present coverage to disable retries is conservative, Auto Remediation presents an amazing potential to finally convey rather more price financial savings with out affecting the person expertise.

Auto Remediation is our first step in leveraging knowledge insights and Machine Studying (ML) for bettering person expertise, lowering the operational burden, and bettering price effectivity of the information platform. It focuses on automating the remediation of failed jobs, but additionally paves the trail to automate operations aside from error dealing with.

One of many initiatives we’re taking, referred to as Proper Sizing, is to reconfigure scheduled large knowledge jobs to request the right sources for job execution. For instance, we’ve famous that the common requested executor reminiscence of Spark jobs is about 4 instances their max used reminiscence, indicating a big overprovision. Along with the configurations of the job itself, the useful resource overprovision of the container that’s requested to execute the job can be decreased for price financial savings. With heuristic- and ML-based strategies, we are able to infer the right configurations of job execution to reduce useful resource overprovisions and save thousands and thousands of {dollars} per yr with out affecting the efficiency. Just like Auto Remediation, these configurations will be mechanically utilized by way of ConfigService with out human intervention. Proper Sizing is in progress and will probably be lined with extra particulars in a devoted technical weblog submit later. Keep tuned.

Auto Remediation is a joint work of the engineers from totally different groups and organizations. This work would haven’t been doable with out the stable, in-depth collaborations. We wish to admire all of us, together with Spark consultants, knowledge scientists, ML engineers, the scheduler and job orchestrator engineers, knowledge engineers, and assist engineers, for sharing the context and offering constructive ideas and precious suggestions (e.g., John Zhuge, Jun He, Holden Karau, Samarth Jain, Julian Jaffe, Batul Shajapurwala, Michael Sachs, Faisal Siddiqi).