druid相关的时间序列数据库——也用到了倒排相关的优化技术

Cattell [6] maintains a great summary about existing Scalable SQL and NoSQL data stores. Hu [18] contributed another great summary for streaming databases. Druid feature-wise sits some-

where between Google’s Dremel [28] and PowerDrill [17]. Druid has most of the features implemented in Dremel (Dremel handles arbitrary nested data structures while Druid only allows for a single

level of array-based nesting) and many of the interesting compression algorithms mentioned in PowerDrill. Although Druid builds on many of the same principles as other distributed columnar data stores [15], many of these data stores are 

designed to be more generic key-value stores [23] and do not sup

port computation directly in the storage layer. There are also other 

data stores designed for some of the same data warehousing issues 

that Druid is meant to solve. These systems include in-memory 

databases such as SAP’s HANA [14] and VoltDB [43]. These data 

stores lack Druid’slowlatency ingestion characteristics. Druidalso 

has native analytical features baked in, similar to ParAccel [34], 

however, Druid allows system wide rolling software updates with 

no downtime. 

Druid is similiar to C-Store [38] and LazyBase [8] in that it has 

twosubsystems,aread-optimizedsubsysteminthehistoricalnodes 

andawrite-optimizedsubsysteminreal-timenodes. Real-timenodes 

are designed to ingest a high volume of append heavy data, and do 

not support data updates. Unlike the two aforementioned systems, 

Druid is meant for OLAP transactions and not OLTP transactions. 

Druid’s low latency data ingestion features share some similar-

ities with Trident/Storm [27] and Spark Streaming [45], however,

both systems are focused on stream processing whereas Druid is 

focused on ingestion and aggregation. Stream processors are great 

complements to Druid as a means of pre-processing the data before 

the data enters Druid. 

There are a class of systems that specialize in queries on top of

cluster computing frameworks. Shark [13] is such a system for 

queriesontopofSpark,andCloudera’sImpala[9]isanothersystem 

focused on optimizing query performance on top of HDFS. Druid

historical nodes download data locally and only work with native 

Druid indexes. We believe this setup allows for faster query laten

cies. 

Druid leverages a unique combination of algorithms in its archi-

tecture. Although we believe no other data store has the same set 

of functionality as Druid, some of Druid’s optimization techniques 

suchas using inverted indices to perform fast filter sarealsousedin

other data stores [26].

 

druid白皮书：http://static.druid.io/docs/druid.pdf