Database Reference
In-Depth Information
named Daniel who work or used to work at LinkedIn. This presearch pro-
cessing dramatically reduces the set of search results by eliminating irrel-
evant results, and allows scoring to focus on more subtle differences—such
as favoring the people with whom the searcher has stronger professional
connections.
Meanwhile, we've been tacking a variety of problems in scoring and ranking.
We've been building personalized, machine-learned ranking models that we
segment by query type. For example, a networker looking up a new acquain-
tance by name is different than a recruiter trying to source candidates. We've
also started using a new architecture called Galene that we built to address
LinkedIn's unique search challenges.
Gutierrez:
What types of queries does Galene now allow LinkedIn users to
do that they couldn't do before?
Tunkelang:
LinkedIn built our early search engines on Lucene, a popular
open-source framework. As we grew, we evolved the search stack by add-
ing layers on top of this framework. Our approach to scaling the system was
reactive, often narrowly focused, and led to stacking new components to our
architecture, each to solve a particular problem without thinking holistically
about the overall system needs. This incremental evolution eventually hit a
wall requiring us to spend a lot of time keeping systems running, and perform-
ing scalability hacks to stretch the limits of the system.
So we decided to completely redesign our platform. The result was Galene, a
new search architecture that is now powering a variety of our search products,
including the "instant" to find people as you type. Galene has also helped us
improve our development culture and processes. For example, the ability to
build new indices every week with changes in offline algorithms supports a
more agile testing and release process.
Gutierrez:
How has the approach evolved since you started the team?
Tunkelang:
When we started the team, we were pretty conservative with
respect to the search experience. Filtering queries by removing irrelevant
results was a pretty radical idea, when conventional wisdom was that you
should return everything and rely on ranking.
Our successes have emboldened us since then. Now we're coming up with
structured query suggestions as searchers type. Our ultimate goal is a
“things-not-strings” experience, where all queries are composed of standard-
ized entities.
