Paper Review: Dremel: Interactive Analysis of Web-Scale Datasets

Paper Review:


Dremel is an interactive ad hoc query system designed to handle real-time web-scale queries with low latency. It uses tree-structured servers and columnar storage with replication to achieve great performance over MapReduce.

Strong points:

  1. Columnar storage is better for faster column retrieval and feature extraction. It exposes data in a more feature-oriented way and make programs easy to process data column by column with better data locality.
  2. Tree structure of server is great. There are more than two layers of servers, which is pretty much all we’ve seen so far (one for metadata and one for data). This structure reminds me of the biggest query system on earth: DNS. With a intermediate layer, the system benefits from one more layers of cache and way better scalabilty. Also it has the potential to yet expand to a bigger scale with another intermediate layer.
  3. The performance is great comparing to similar systems like Hive, which is not real time. It delivers all the requirements as a real-time interactive ad hoc querying system. However, it seems to me that Apache Drill could achieve pretty much the same thing with flexibility on data types.

Weak points:

  1. Relatively poor performance when few columns are read or dealing with unstructured data. In that case we cannot take advantage of the columnar storage. But I guess they are pretty sure about the query types that Dremel is going to handle so it’s fine. Dremel is design to deal with the Ad Hoc queries of structured data ready to be analysed.
  2. I don’t think MapReduce and Dremel make a valid comparison. Of course users can still use MapReduce do perform Dremel’s query and analysis job but that’s not what MapReduce is designed for, which is distributed batch processing. Those two are more complimentary rather than comparable to me, and that’s exactly what the authors suggested in the observation in the paper: “MR and query processing can be used in a complementary fashion; one layer’s output can feed another’s input”.
  3. There’s not you can do to modify the data (update or creation) expect append, which limits what users could perform with Dremel. I guess implement update method is not in their development priority since the data analysis rarely used modification but still it’s nice to have a flexible way to change the data.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s