April 15, 2024
Clark Wright
The Airbnb Tech Blog

By: Clark Wright

Nowadays, as the amount of knowledge collected by corporations grows exponentially, we’re all realizing that extra information isn’t at all times higher. The truth is, extra information, particularly should you can’t depend on its high quality, can hinder an organization by slowing down decision-making or inflicting poor selections.

With 1.4 billion cumulative visitor arrivals as of year-end 2022, Airbnb’s progress pushed us to an inflection level the place diminishing information high quality started to hinder our information practitioners. Weekly metric experiences had been troublesome to land on time. Seemingly primary metrics like “Energetic Listings” relied on an internet of upstream dependencies. Conducting significant information work required vital institutional data to beat hidden caveats in our information.

To satisfy this problem, we launched the “Midas” course of to certify our information. Beginning in 2020, the Midas course of, together with the work to re-architect our most crucial information fashions, has introduced a dramatic improve in information high quality and timeliness to Airbnb’s most crucial information. Nonetheless, reaching the complete information high quality standards required by Midas calls for vital cross-functional funding to design, develop, validate, and preserve the mandatory information belongings and documentation.

Whereas this made sense for our most crucial information, pursuing such rigorous requirements at scale offered challenges. We had been approaching some extent of diminishing returns on our information high quality investments. We had licensed our most crucial belongings, restoring their trustworthiness. Nonetheless, for all of our uncertified information, which remained nearly all of our offline information, we lacked visibility into its high quality and didn’t have clear mechanisms for up-leveling it.

How might we scale the hard-fought wins and greatest practices of Midas throughout our whole information warehouse?

On this weblog put up, we share our revolutionary strategy to scoring information high quality, Airbnb’s Information High quality Rating (“DQ Rating”). We’ll cowl how we developed the DQ Rating, the way it’s getting used right this moment, and the way it will energy the subsequent chapter of knowledge high quality at Airbnb.

In 2022, we started exploring concepts for scaling information high quality past Midas certification. Information producers had been requesting a lighter-weight course of that would present among the high quality guardrails of Midas, however with much less rigor and time funding. In the meantime, information customers continued to fly blind on all information that wasn’t Midas-certified. The model round Midas-certified information was so robust that buyers began to query whether or not they need to belief any uncertified information. Hesitant to dilute the Midas branding, we wished to keep away from introducing a light-weight model of certification that additional stratified our information with out actually unlocking long-term scalability.

Contemplating these challenges, we determined to shift to an information high quality technique that pushed the incentives round information high quality on to information producers and customers. We made the choice that we might now not depend on enforcement to scale information high quality at Airbnb, and we as a substitute wanted to depend on incentivization of each the info producer and client.

To completely allow this incentivization strategy, we believed it could be paramount to introduce the idea of an information high quality rating immediately tied to information belongings.

We recognized the next targets for the rating:

  • Evolve our understanding of knowledge high quality past a easy binary definition (licensed vs uncertified).
  • Align on the enter elements for assessing information high quality.
  • Allow full visibility into the standard of our offline information warehouse and particular person information belongings. This visibility ought to 1) Create pure incentives for producers to enhance the standard of the info they personal, and a pair of) Drive demand for high-quality information from information customers and allow customers to determine if the standard is suitable for his or her wants.

Earlier than diving into the nuances of measuring information high quality, we drove alignment on the imaginative and prescient by defining our DQ Rating guiding rules. With the enter of a cross-functional group of knowledge practitioners, we aligned on these guiding rules:

  • Full protection — rating could be utilized to any in-scope information warehouse information asset
  • Automated — assortment of inputs that decide the rating is 100% automated
  • Actionable — rating is simple to find and actionable for each producers and customers
  • Multi-dimensional — rating could be decomposed into pillars of knowledge high quality
  • Evolvable — scoring standards and their definitions can change over time

Whereas they might appear easy or apparent, establishing these rules was vital as they guided every resolution made in growing the rating. Questions that in any other case would have derailed progress had been mapped again to our rules.

For instance, our rules had been vital in figuring out which objects from our wishlist of scoring standards needs to be thought-about. There have been a number of inputs that definitely might assist us measure high quality, but when they may not be mechanically measured (Automated), or in the event that they had been so convoluted that information practitioners wouldn’t perceive what the criterion meant or the way it could possibly be improved upon (Actionable), then they had been discarded.

We additionally had a set of enter alerts that extra immediately measure high quality (Midas certification, information validation, bugs, SLAs, automated DQ checks, and so on.), whereas others had been extra like proxies for high quality (e.g., legitimate possession, good governance hygiene, the usage of paved path tooling). Had been the extra specific and direct measurements of high quality extra invaluable than the proxies?

Guided by our rules, we ultimately settled on having 4 dimensions of knowledge high quality: Accuracy, Reliability (Timeliness), Stewardship, and Usability. There have been a number of different potential dimensions that we thought-about, however these 4 dimensions had been probably the most significant and helpful to our information practitioners, and made sense as axes of enchancment, the place we care and are keen to spend money on enhancing our information alongside these dimensions.

Every dimension might combine implicit and specific high quality indicators, with the important thing being: Not each information client wants to totally perceive each particular person scoring element, however they’ll perceive {that a} dataset that scores poorly on Reliability and Usability struggles with touchdown on-time persistently and is troublesome to make use of.

We might additionally weigh every dimension in line with our notion of its significance in figuring out high quality. We thought-about 1) what number of scoring elements belonged to every dimension, 2) enabling fast psychological math, and three) which parts our practitioners care about most to allocate 100 complete factors throughout the scale:

The “Dimensions of Information High quality” and their weights

In the meantime, if desired, the scale could possibly be unpacked to get to a extra detailed view of knowledge high quality points. For instance, the Stewardship dimension scores an asset for high quality indicators like whether or not it’s constructed on our paved path information engineering instruments, its governance hygiene, and whether or not it meets legitimate information possession requirements.

Unpacking the Information Stewardship Dimension

We knew surfacing the DQ Rating in an explorable, actionable format was vital to its adoption and success. Moreover, we needed to floor information high quality data immediately within the venue the place information customers already found and explored information.

Fortunately, we had two present instruments that will make this a lot simpler: Dataportal (Airbnb’s information catalog and exploration UI), and the Unified Metadata Service (UMS). The rating itself is computed in a every day offline information pipeline that collects and transforms varied metadata parts from our information methods. The ultimate process of the pipeline uploads the rating for every information asset into UMS. By ingesting the DQ Rating into UMS, we will floor the rating and its elements alongside each information asset in Dataportal, the start line for all information discovery and exploration at Airbnb. All that remained was designing its presentation.

Certainly one of our objectives was to floor the idea of high quality to information practitioners with various experience and desires. Our person base had absolutely adopted the licensed vs uncertified dynamic, however this was the primary time we might be presenting the idea of a spectrum of high quality, in addition to the standards used to outline high quality.

What could be probably the most interpretable model of a DQ Rating? We would have liked to have the ability to current a single information high quality rating that held which means at fast look, whereas additionally making it potential to discover the rating in additional element.

Our closing design presents information high quality in 3 ways, every with a special use case in thoughts:

  1. A single, high-level rating from 0–100. We assigned categorical thresholds of “Poor”, “Okay”, “Good”, and “Nice” based mostly on a profiling evaluation of our information warehouse that examined the present distribution of our DQ rating. Greatest for fast, high-level evaluation of a dataset’s total high quality.
  2. Dimensional scores, the place an asset can rating completely on Accuracy however low on Reliability. Helpful when a specific space of deficiency isn’t problematic (e.g., the buyer needs the info to be very correct however isn’t frightened about it touchdown shortly every single day).
  3. Full rating element + Steps to enhance, the place information customers can see precisely the place an asset falls quick and information producers can take motion to enhance an asset’s high quality.

All three of those shows are proven within the screenshots under. The default presentation gives the dimensional scores “Scores per class”, the explicit descriptor of “Poor” together with the 40 factors, and steps to enhance.

Full information high quality rating web page in Dataportal

If a person explores the complete rating particulars, they will study the precise high quality shortcomings and examine informative tooltips offering extra element on the scoring element’s definition and benefit.

Full rating element presentation

For information producers, the rating is offering

  • Clear, actionable steps to enhance the DQ of their belongings
  • Quantified DQ, measuring their work
  • Clear expectations round DQ
  • Targets for tech debt clean-up

For information customers, the DQ Rating

  • Improves information discoverability
  • Serves as a sign of trustworthiness for information (identical to how the evaluate system works for Airbnb Friends and Hosts)
  • Informs customers of the precise high quality shortcomings to allow them to be comfy how they’re utilizing the info
  • Permits customers to hunt out and demand information high quality

From a information technique perspective, we’re leveraging inside question information mixed with the DQ Rating to drive DQ efforts throughout our information warehouse. By contemplating each the amount and the kind of consumption (e.g., whether or not a specific metric is surfaced in our Govt reporting), we’re capable of direct information groups to probably the most impactful information high quality enhancements. This visibility has been very enlightening for groups who had been unaware of their lengthy tail of low-quality belongings, and has enabled us to double down on high quality investments for heavy-lift information fashions that energy a big share of our information consumption.

Lastly, by growing the DQ Rating, we had been capable of present uniform steering to our information producers on producing high-quality, albeit uncertified belongings. The DQ Rating has not changed certification (e.g., solely Midas-certified information can obtain a DQ Rating > 90). We proceed to certify our most crucial subset of knowledge, and consider the use circumstances for these belongings will at all times benefit the guide validation, rigor, and maintenance of certification. However for every little thing else, the DQ Rating reinforces and scales the rules of Midas throughout our warehouse.

We’re enthusiastic about now with the ability to measure and observe quantified enhancements to our information high quality, however we’re simply getting began. We just lately expanded on the unique DQ Rating to attain our Minerva metrics and dimensions. Equally, we plan to carry the identical idea of a DQ Rating to different information belongings like our occasion logs and ML options.

As the necessities and calls for towards our information proceed to evolve, so will our high quality expectations. We’ll proceed to evolve how we outline and measure high quality, and with fast enchancment in areas like metadata administration and information classification, we anticipate additional effectivity and productiveness positive factors for all information practitioners at Airbnb.

The DQ Rating wouldn’t have been potential with out a number of cross-functional and cross-org collaborators. They embody, however should not restricted to: Alvin Wo, Gang Feng, Mark Steinbrick, Chitta Shirolkar, Jonathan Parks, Sylvia Tomiyama, Felix Ouk, Jason Flittner, Ying Pan, Logan George, Woody Zhou, Michelle Thomas, and Erik Ritter.

Particular because of the broader Airbnb information group members who supplied enter or help to the implementation workforce all through the design, improvement, and launch phases.

If such a work pursuits you, take a look at a few of our related positions.


All product names, logos, and types are property of their respective house owners. All firm, product and repair names used on this web site are for identification functions solely. Use of those names, logos, and types doesn’t suggest endorsement.