Skip to main content

12 posts tagged with "wayfair"

View All Tags

· 5 min read

Happened Before

This is our origin story. The simplified high level architecture of our logging systems at the time of tremor’s birth is:

old pipeline

Identified Need

Enable load shedding and consistent capacity management at production scale of the logging and metrics firehoses within Wayfair’s production, staging and development environments; even when our live production working set exceeds planned capacity, but especially when catastrophic failure of infrastructure overwhelms our storage tier.

Required Outcome

Our network operations center, system reliability and service organization engineers should never be in a position where in-flight mission-critical events are unavailable or lagging ( seconds or even minutes ) the live working set state of our production systems.

Stale operational events direct limited resources under time constraints and production service level objective pressures to curing or tending to the wrong problems, at the wrong times and lead to missed opportunities for early and cheap resolution.


It is understood that, if a critical system ( such as a production database ) fails - query failures won’t have a single point of origin in the logs - whether over time, or across application or business units. When under extreme load beyond capacity - only a subset of the firehose is required for in-flight diagnostics of our production estate.

In-flight categorization, classification & rate limiting enables just-in-time traffic shaping, load-shedding and bending critical infrastructure to available capacity even as capacity changes in planned or surprising ways.

It is understood that, regardless of the system state - that some logs from a subset of systems, applications or services are relatively more important than others. This is even more true when our systems and services are under extraordinary strain due to unplanned failures and outages. Unplanned failures and outages are expected at Wayfair production scale.

The relative priority of classified and categorized events is known and the relative priority is subject to change over time, as are the event classifications and categorizations.

It is likely impossible to convince thousands of engineers in hundreds of teams across our technology organization to stop using JSON format in the short, medium and possibly even longer term.

Snot. It’s a JSON rock’n’roll world. I guess we’ll just have to deal with it


This works fine 99% of the time - but under extreme load it just does not work at our scale and the impact to our network operations and service reliability and technical communities is severe at the exact moments when we are at peak trading periods and impact to the business is highest - a perfect storm.

old pipeline

In our ELK ( ElasticSearch, Logstash, Kibana ) based logging and Influx ( Telegraf, Chronograf, InfluxDB ) based metrics environments there are limited means to dynamically tune the production system to accommodate highly variant working sets. As we are an online eCommerce system we often go through sustained periods of nominal activity followed by insane periods of extreme peak activity - often hitting peaks we’ve never before experienced.

We do however, have a very experienced infrastructure and platform engineering organization that are deeply attuned to our working set and changing business and operational needs.

The initial version of tremor simply preserved the overall pipeline:

  • Applications in a number of programming languages generate logs
  • A Logstash sidecar receives, normalizes and forwards logs to elasticsearch
  • ElasticSearch indexes logs for display by SRE’s and developers in Kibana

The processing stages remained unchanged ( and largely serviced by Logstash ):

  • Enrich raw logs with business unit, organization, application and environment metadata
  • Normalize schemas - map logs to a common structure and well known field names
  • Known unknowns - isolate user defined fields and extensions to a well known slush field

But added an extra processing stage, a tremor cluster that added back-pressure detection and load-adaptive rate-limiting based on in-flight analysis of the business events in the firehose based on real-time categorization and classification of those event streams.

Now, in the ( significantly less than 1% ) of time when our systems are over-capacity or experiencing catastrophic unplanned events - tremor can discard events based rate limits and event classifications that can be tuned by our SRE, NOC and systems architects and developers.


Tremor’s core mission and mandate started out with, and continues to be the non-functional quality of service of mission critical at scale event based systems. Providing system operators with first class means to shape and tune capacity, load and production countermeasures of the production firehose without compromising the business with system downtime is the core requirement.

The most effective way to achieve this in a highly fluid and ever changing environment is what has yielded the tremor project as it became to be - a general purpose event processing system designed for at scale distributed mission critical systems.

When at or over capacity systems no longer bend to the business - tremor is how we inject a little elasticity into our production quality of service - enriching our systems and services with just-in-time load-adaptive bendability.

new pipeline

· 7 min read


Development on Tremor started in the autumn of 2018 as a point solution for introducing a configurable traffic-shaping mechanism for our infrastructure platform engineering organization.

There were two use cases being considered at the time- the traffic-shaping use case, which had the highest priority; and a distributed data replication use case for our Kafka-based environments.

The Origin of Tremor

The Traffic Shaping use case required that log information streaming from various systems be categorized and classified in-flight, so that temporally bound rate limits could be applied when our production estate has a working set that is in excess of our current capacity. As a 24x7x365 eCommerce environment with year-on-year higher peaks, higher volumetrics and higher loads on our systems, this is a frequent occurrence. Tremor was designed to remove the burden of scaling these systems from our system reliability and network operations center engineers.

This use case saw the birth of Tremor, which was delivered in 6 weeks, just in time for Cyber 5 (the long bank-holiday weekend in the US centered around Thanksgiving, where US folk go peak shopping!). At the time, Tremor was inflexible; only supported connectivity to Kafka for Ingres and Elasticsearch downstream; and, the classification and rate-limiting logic was more or less hardwired- but it worked flawlessly!

Although designed with our logging systems in mind, our metrics systems team saw that Tremor would work for their environment (InfluxDB downstream rather than Elasticsearch, but the logic was basically the same). Tremor was designed to detect backpressure on downstream systems, and to intelligently- based on user defined logic- react by adaptively tuning the data distribution to selectively discard or forward data based on a classification system and rate limits.

The Rise of Event Processing

The user-defined logic required by our logging and metrics teams very quickly became sophisticated. A fairly quick succession of releases saw the logic change from fairly simple filtering, classification (enrichment) and rate limiting (selective dropping due to backpressure) to richer needs to slice and dice nested JSON-like data, to normalize these datasets. This rising need for sophistication in processing along with the adoption of Tremor to replace Logstash and other log shipping frameworks in our source nodes, and to replace Telegraf in our metrics environments, gave birth to Tremor as an event processing engine in its own right.

With richer programming primitives, Tremor advanced from traffic shaping, and log and metrics shipping to log and metrics cleansing, normalisation, enrichment and transformation, to impose a common symbology and structure on logs and metrics generated from many different programming languages, platforms and tools across our production estate- from system logs originated via Syslog, to those via GELF, or Prometheus.

The YAML-based configuration of the query pipelines quickly became error-prone and cumbersome as the sophistication, complexity and size of application logic grew. In addition, the success of our scripting language, tremor-script, and its hygienic set of tooling and IDE integration encouraged the addition a query language, thus enabling data flow processing with Tremor.

Cloud Native Migration

At the same time, our technology organisation was preparing for migration from on-premise data centres to cloud-based systems. Our legacy VM-based systems were now being containerised and deployed via Kubernetes in our on-premise environments. Simultaneously, our Cloud-native infrastructure, configuration as code, secrets management, developer scaling and many other infrastructure and development platform teams were building out the services that would allow our line of business service engineers to migrate to the cloud.

Tremor was already battle-hardened as a sidecar deployment, and this was extended to Cloud-native deployments by our kubernetes team who packaged Tremor for kubernetes as a set of helm charts.

As a relatively fast-paced technology organisation, our logging and metrics teams weren’t stalled- the scripting and query languages enabled rapid development. However, lack of native support in Tremor to modularise and reuse logic became a limiting factor. Thus, over the span of approximately a year, most of the modularity mechanisms, now standard in Tremor, were developed.

Tremor was now battle-tested, battle-hardened and widely adopted, having gone through many enhancements and replacing many disparate tools and frameworks with a single, easy-to-operate solution.

Tremor is the first major component of in-house critical infrastructure that was open-sourced by Wayfair and contributed to the Linux Foundation under the Cloud Native Computing Foundation.

The Tremor team then selected a panel of experts and friends, who had early access to Tremor, to assist with the open-sourcing process.

Tremor Today

Today, Tremor is planned and progresses in the open. Anyone can contribute, and with contribution comes maintainership. Tremor is now a CNCF community (sandbox) project, and maintainership has grown beyond Wayfair.

At Wayfair, Tremor has been adopted for our search environments. Our search domain has some complex use cases for audited document elementisation and indexing that are multi-participant and that need to occur transactionally.

Tremor’s QoS mechanisms were extended to support transaction orchestration to enable this and similar use cases. Our search teams are also leveraging Tremor for its traditional areas of strength in traffic shaping and adaptive rate limiting.

Our logging and metrics teams are now recalibrating our technology organisation's philosophy and core infrastructure that supports observability across our production estate. The rise of CNCF OpenTelemetry offers developers a consistent and stable set of primitives for logs, metrics and distributed tracing. It’s a lingua franca that offers developers consistency.

Even more important to large or extreme scale organisations is flexibility to make infrastructure and service decisions based on capacity, cost, or other concerns. Through CNCF OpenTelemetry as a core building block, now natively supported in Tremor, our observability teams can now unify our production support, operations and services around Tremor-based OpenTelemetry- preserving the key values that Tremor adds, whilst opening up the Cloud Native new possibilities that OpenTelemetry and OpenTelemetry-based services offer.

Tremor Tomorrow

As the tremor community grows, we have seen great contributions from the growing tremor community adding support for many different types of systems - ranging from the addition of AMQP support by Nokia Communications through to our ongoing collaboration with Microsoft Research on the snmalloc allocator which tremor defaults to.

The latest production solution based on tremor at Wayfair is from our Developer Platforms technology organization - who have used an extension to PHP to identify dead code in our legacy PHP monolith. This solution is based on tremor and included contributions to tremor in the form of the unix_socket connectivity. Thank you Ramona!

The Tremor maintainers also work on the Rust port of simd-json- SIMD accelerated JSON, which is a sister project to Tremor. Mentorships via the Linux Foundation are also producing some wonderful new capabilities and enhancements to the system, ranging from better support for developing gRPC-based connectors in Tremor, to GCP connectivity, or work to add a plugin development kit to Tremor.

Wayfair is already benefiting from community interest and contributions to the project - and the Tremor team has hired its newest member as a result of open-sourcing the project. We are hoping to package and open-source more components of Wayfair’s Tremor-based systems through the Tremor Project, and through other open-source projects at Wayfair through our new Open Source Program Office.

If you are interested, hop on over to our community chat server and say hello!