How to Build a Google Search Autocomplete

Whenever you start typing your search on Google, you get a list of recommendations, and the more letters you type, the more accurate the recommendations get. If you're like me, you've always wondered how this works — is the inverted index being stored, or is it something else?

The data structure that would be apt here is a Trie.

You may also like: Implementing Low-Level Trie: Solving With C++.

System Requirements

Considering the scale of Google, the factors that we need to keep in mind are latency, consistency, and availability. A desirable latency should be very low, giving/changing suggestions on each letter you would type. Next, the system needs to be available all the time; however, the consistency can be comprised here. Each time you type something, it would change the frequency of the previously stored query, which would affect the suggestions. The slight delay here is okay and eventual consistency would also work.

The Concept:

Approach #1

A Trie represents a word as a tree with each letter as a node and the next letter at its child node and so on. Google also stores each word/sentence in the form of a trie. Consider here, the parent node is “h,” its child node is “a,” then “r” and so on. Each node can have 26 child nodes. Now, each node can also store the frequency of each letter searched.

For example, node “h” stores the search frequency of “h." Its child node, “a,” stores the search frequency of “ha” and so on. Now, if we want to show the top N recommendations, say you typed “h,” and the suggestions should show “harry potter” or “harry styles.” Then, we need to sort all recommendations from the parent node down to every level on the frequency and show it. This would mean scanning terabytes of data, and as latency is our goal, this scanning approach would not work. 

Approach #2

To make this approach more efficient, we can store more data on each node, along with the search frequency. Lets store the top N queries at each node from the subtree below it. This means that the node “h” would have queries like “harry potter,” “harley davidson,” etc stored. If you traversed down the tree to “harl” (i.e. you type “harl”), the node, “l,” would have queries like “harley davidson,” “harley quinn,” etc.

This approach is better compared to the previous one, as the read is quite efficient. Anytime a node’s frequency gets updated, we traverse back from the node to its parent until we reach the root. For every parent, we check if the current query is part of the top N. If so, we replace the corresponding frequency with the updated frequency. If not, we check if the current query’s frequency is high enough to be a part of the top N.

If so, we update the top N with the frequency. Though this approach works, it does affect our read efficiency — we need to put a lock on the node each time we do write/update, so that the user won’t get stale values, but if we consider eventual consistency, then this might not be much of an issue. The user might get stale values for a while, but eventually, it would get consistent. Still, we will look at an extension of this approach.

Approach #3

As an extension of the previous approach, we can store the data offline. We can store a hashmap of a query to its frequency, and once the frequency reaches a set threshold value, we can then map it to the servers.

Scaling

Now, there wouldn’t be just one big server that stores all petabytes of data; we would vertically scaling for life — there is a better approach. We can distribute data (sharding) by prefixes on various servers. For example, prefixes like "a," "aa," "aab," etc. would go on server #1 and so on. We could use a load balancer to keep the map of the prefix with the server number.

But consider this, some servers with data like "x,” “xa,” “yy,” etc. would have less traffic compared to the letter “a.” So, there can be a threshold check on each server; if the load surpasses that traffic, then the data can be distributed among other shards.

If you are concerned about a single point of failure, there can be many servers acting as load balancer, so if any load balancer goes down, it is replaced by another one. You can use ZooKeepers to continuously health check load balancers and act accordingly.

Further Reading

• ZooKeeper: A Real World Example of How to Use It.
• Apache Solr: Getting Optimal Search Results.
• Kafka Producer and Consumer Examples Using Java.