really loved how you were suggesting solutions and evaluating them. Most solutions out there just touch a naive sql query solution and jump directly to a trie.
At 12:55 you should push onto the heap before you pop, otherwise if the value you push is smaller than the value you are popping, your result will be incorrect.
Question: 40:46 Why do we need something like Flink to consume the messages from Kafka, if we are already sharding (in Kafka) the incoming data based on the search term range? Can you also clarify why we need an aggregation service in front of Kafka? Overall a great video, Jordan. Thank you!
You always need something to consume kafka messages, they can't just go straight to HDFS. In this case, we're just using flink to store messages in memory until we get enough of them, flush them to a parquet file, and then send to hadoop. In theory, we could also do some pre-aggregations/counting in flink too if we're really optimizing.
Great video Jordan. Had this question to ask: How does the design take care of server going down that was handling "app" (from word apple) partition. Trie data stored on that server would be lost so I believe data would be replicated across multiple nodes? Also, how would the switchover take place in this scenario? Zookeeper listening to all the active/replica nodes and switching over when it stops getting heartbeat? Is that correct understanding?
Yes, we'd want all nodes to have replicas. Zookeeper can listen to them and round robin requests to these nodes. In this case, all nodes are replicas as we don't write to any of them, they're all read only.
Question: @jordan @41:21, it is unclear to me how our queries are processed with the trie distributed across multiple servers. Lets say there are only two servers, one contains trie for words [a, b, aa,... abc] and other for [abcd,... zzzz]. In this case when I start typin and have typed "a", then I get connected to first server via websocket, and then when I type b and c, I continue to be connected and traverse the trie down to character 'c' on first server. Now, when I type 'd' what happens? Is it correct that: a) The first server (to which I am connected) will see that 'c' has no children, thus will break the websocket connection, therefore b) client will reconnect saying "gimme stuff for abcd", and load balancer directs me to second server, and c) the second server's trie has root node of "abc", so it goes down one step to 'd' and return the values? Is my understanding correct above?
I'd think that at the node at which the partition begins on system 1 it would say something like "no longer on this system" so that the client knows it needs to reach back out to the load balancer.
34:04 I don't see how the server has to be stateful. If I'm at "ap" and type in another "p", the server doesn't need to know that I was at "ap" previously, because now "app" is a new prefix query. So the server can be stateless. Is this not correct?
@ 21:08 short is 2bytes so its 16bits not 8 and hence we have like 65k terms ( i have to point this minor insignificant mistake or i cannot go to bed since I'm a internet police)
Hey Jordan, thanks for this video! I'm curious about how trie can be stored in different databases, could you share something related? Additionally, I think we can have cache in front of suggestion service, or probably what you mean by saying suggestion service already include that? Thanks!
I guess it can't really, hence our issue. You can be cheeky and use the tactic I used to store it in spark, however then you lose some of the nice time complexity As for Q2, the suggestion service is effectively already a cache - it's in memory and has cached the top suggestions for each search term
@@jordanhasnolife5163 I'm not sure I agree. Elasticsearch supports term and phrase suggestions as special use cases, and it gives users control over general relevance features. I work on a search team, and our design for this feature is centered around an Elasticsearch cluster w/ special typeahead indices, an ETL from BQ to that cluster, and a service to query the cluster. I don't know if our design is the industry standard, and it depends on exactly what you're trying to do, but I think this is definitely one of the ES use cases. (Typeahead isn't just about popularity either, there could be many different heuristics you need to use to rate which suggestions are the best. There may be machine learning models involved to help determine that as well.)
Another great video! Thanks for making it. I am a bit confused about the update path. 1. It looks like we are creating new trie from the logs (containing search term with freq in kafka) instead of updating the existing trie. Lets say we want to account for last few days of search, then to build the trie shouldn't we feed the copy of existing trie as well (along with recent search logs) to hdfs to calculate top suggestions for each prefix? 2. Instead of app server just getting data about top suggestion for each prefix from hdfs, is it possible for us to compute the trie as well offline and then load it in server? If yes, can you also please suggest tools to use for computing trie offline and loading from offline to server memory ?
1) HDFS already has the last few days of data available. It doesn't have to delete that just because we computed another trie from it. You wouldn't have to send the existing trie. 2) Considering that you can't really represent a trie in a text file like that, I'm not quite sure. I guess in theory, you could compute it on one server from the hdfs data, then serialize it to JSON or something, then send it out to all of the other servers. But then even, you're just building a trie from the JSON rather than the frequencies which frankly has a similar time complexity.
Our servers are inherently not stateless. They need to keep a pointer to their current location in the trie, or else we won't have constant time operations when a user types an additional character.
The calculation of runtime at 13:40 is wrong. Runtime is O(log(n) + m) where is n is total words in dictionary, and m is total words matched by the query
I assume you're using this to account for the actual computation cost of finding the start and end of where we're reading. Good point. That being said, for the m piece I think it would be m * log(k), where k is the number of typeahead suggestions to return, as we need to formulate our heap.
I think the main point is that HDFS is designed for processing large sequential data write like GB level, not good for small data write/read, thus message queue is better for keep event with small data size
A follow up video with a few more concepts would be great 1. We may not always need to reconnect a new web socket. That layer can be decoupled with suggestion service 2. How do we handle cases where updating the trie without impacting read performance- I’d guess something like update replica async and point traffic to it when ready? 3. What can be cached on client vs always query backend 4. Personalization / contextual results
@@jordanhasnolife5163 it's more good idea to create a video and just tagging the title as "Type Ahead Suggestion" and then not explaining any core concepts related to "Type Ahead Suggestion" and just going over few simple tree diagrams.
@@lv0320 1) agree 2) agree, store a second copy and switch over when done 3) pretty much anything and everything can hopefully be cached on the client, which speaks to 4) we can perhaps do some of that personalization piece locally (in the case of typeahead, likely tougher if we wanted to tailor search results)