Kroger buys into Google Cloud migration, despite challenges
It wasn't easy for Kroger to migrate its customer-facing search engine to Google Cloud, but it happened fast, thanks to proper planning and the right mix of talent.
SAN FRANCISCO -- The Kroger Co. just might make more shopping trips to Google's cloud platform after a successful cloud migration of its website and mobile app's search engine.
The retailer ran its search engine on premises with a tangle of technologies -- among them, Spring, Kafka, Solr and Cassandra, said Jeff Hecht, engineering manager for the company's digital search team. Hecht spoke at the Google Cloud Next conference here last week.
Kroger is the United States' biggest supermarket chain in terms of revenue, with $121.2 billion in fiscal 2018, and it has 2,800 stores that serve 14 million shoppers each day. Its search engine helps customers shop, but also generates valuable insight into their behaviors that can help Kroger predict demand for products.
Kroger's annual digital revenue run rate grew 58% last year to more than $5 billion, as more shoppers purchase goods online, rather than in physical stores. The company expects digital sales to top $9 billion this year. To keep that momentum, Kroger's search engine must constantly evolve and improve.
"The big question we had was, 'Do we do a lift-and-shift and move it out [to the cloud], or do something to completely replatform the search solution we had?'" Hecht said.
Then, there was the question of how to start the cloud migration project. Kroger's internal staff had limited experience with cloud development projects, Hecht said. Kroger also had to keep the existing search engine running while it made the move. Ultimately, the supermarket giant assembled a team that consisted of seven in-house engineers -- six from Google and five from a partner, SpringML.
The team designed a cloud infrastructure reference architecture as part of the search engine project. Everything is done via infrastructure as code in the interest of speed and flexibility, Hecht said.
Still, Kroger had somewhat of a learning curve as it moved to Google Cloud Platform (GCP).
When you're doing your discovery phase [for a cloud project], think about the stuff you take for granted.
Jeff HechtEngineering manager for the digital search team, Kroger
"Network architectures [in the cloud] are hard, especially when you're used to having your own data center," he said.
The company also had to preserve the blue-green methodology to deploy containers both on premises and on GCP with Google Kubernetes Engine. In a blue-green deployment, two identical production environments sit side by side, with one idle. Updates and changes are made to the idle environment and fully tested; once this work is done, the idle environment goes live in the place of the other one.
Hecht also offered a general tip to the session's attendees: "When you're doing your discovery phase [for a cloud project], think about the stuff you take for granted," he said.
Ultimately, Kroger's project produced a supportable environment on GCP within about six months, despite the fact its team completely rewrote the search engine in the process, Hecht said. Searches now run eight times faster, and the underlying infrastructure footprint is smaller than the previous on-premises installation, he added.
Kroger chose Google, and not AWS, for a simple reason that's become common among retailers that want to put resources into the cloud: "We don't want to put money in the coffers of our competitor," Hecht said. Amazon's retail division competes with Kroger online and through brick-and-mortar locations with Whole Foods Market.
But Kroger isn't all-in with GCP. In fact, it recently partnered with Microsoft to co-develop a retail software platform on Azure and a connected store experience driven by Azure IoT. Hecht's team isn't involved with that effort, he said in an interview after the presentation.
Google's reputation in search made it a natural pick for Kroger's project, Hecht said.
And that's even more relevant as search evolves beyond term frequency and best matches and becomes more of a data science problem.
"If you're doing product search, that's not going to cut it," Hecht said. "You have to understand what AI, what ML [machine learning], what types of things am I going to do to enhance that search result for the customer. That's huge with Google."
