Monday, August 12, 2019
Most of the time, our search engine runs properly. Our teams work hard to prevent technical issues that could affect our users who are searching the web, or webmasters whose sites we index and serve to users. Similarly, the underlying systems that we use to power the search engine also run as intended most of the time. When small disruptions happen, they are largely not visible to anyone except our teams who ensure that our products are up and running. However, like all complex systems, sometimes larger outages can occur, which may lead to disruptions for both users and website creators.
In the last few months, such a situation occurred with our indexing systems, which had a ripple effect on some other parts of our infrastructure. While we worked as quickly as possible to remedy the situation, we apologize for the disruption, as our goal is to continuously provide high-quality products to our users and to the web ecosystem.
Since then, we took a closer, careful look into the situation. In the process, we learned a few lessons that we'd like to share with you today. In this blog post, we will go into more details about what happened, clarify how we plan to communicate better if such things happen in the future, and remind website owners of the channels they can use to communicate with us.
So, what happened a few months ago?
In April, we had several issues related to our index. The Search index is the database that holds the hundreds of billions of web pages that we crawled on the web and that we think could answer some of our users' queries. When a user enters a query in the Google search engine, our ranking algorithms sort through those pages in our Search index to find the most relevant, useful results in a fraction of a second. Here is more information on what happened.
1. The indexing issue
To start it off, we temporarily lost part of the Search index.
Wait... What? What do you mean "lost part of the index?" Is that even possible?
Basically, when serving search results to users, to accelerate the speed of the service, the query of the user only "travels" as far as the closest of our data centers supporting the Google Search product, from which the Search Engine Results Page (SERP) is generated. So when there are modifications to the composition of the index (some pages added and removed, documents are merged, or other types of data modification), those modifications need to be reflected in all of those data centers. The consequence is that users all over the world are consistently served pages from the most recent version of the index.
Keeping the index unified across all those data centers is a non trivial task. For large user-facing services, we may deploy updates by starting in one data center and expand until all relevant data centers are updated. For sensitive pieces of infrastructure, we may extend a rollout over several days, interleaving them across instances in different geographic regions. source.
So, as we pushed some planned changes to the Search index, on April 5th parts of the deployment system broke, on a Friday no-less! More specifically: as we were updating the index over some of our data centers, a small number of documents ended up being dropped from the index accidentally. Hence: "we lost part of the index."
Luckily, our on-call engineers caught the issue pretty quickly, at the same time as we started picking up chatter on social media (thanks to everyone who notified us over that weekend!). As a result, we were able to start reverting the Search index to its previous stable state in all data centers only a few hours after the issue was uncovered (we keep back-ups of our indexes just in case such events happen).
We communicated on Sunday, April 7th that we were aware of the issue, and that things were starting to get back to normal. As data centers were progressively reverting back to a stable index, we continued updating on Twitter (on April 8th, on April 9th), until we were confident that all data centers were fully back to a complete version of the index on April 11th.
2. The Search Console issue
Search Console is the set of tools and reports any webmaster can use to access data about their website's performance in Search. For example, it shows how many impressions and clicks a website gets in the organic search results every day, or information on what pages of a website are included and excluded from the Search index.
As a consequence of the Search index having the issues we described above, Search Console started to also show inconsistencies. This is because some of the data that surfaces in Search Console originates from the Search index itself:
- the Index Coverage report depends on the Search index being consistent across data centers.
- when we store a page in the Search index, we can annotate the entry with key signals about the page, like the fact that the page contains rich results markup for example. Therefore, an issue with the Search index can have an impact on the Rich Results reports in Search Console.
Basically, many Search Console individual reports read data from a dedicated database. That database is partially built by using information that comes from the Search index. As we had to revert back to a previous version of the Search index, we also had to pause the updating of the Search