druid实时分析数据存储

Druid是一个开源的数据存储系统,用来在大数据集上做实时数据分析和查询,Druid融合了列式存储结构,分布式的独立架构,并利用高级索引结构,可以秒级延时随意查询几十亿行数据,这篇文章描述了Druid结构,实现细节,揭露了它如何在非常低延时下,做到快速聚合,快速过滤和数据处理。
展开查看详情

1. Druid A Real-time Analytical Data Store Fangjin Yang Eric Tschetter Xavier Léauté Metamarkets Group, Inc. echeddar@gmail.com Metamarkets Group, Inc. fangjin@metamarkets.com xavier@metamarkets.com Nelson Ray Gian Merlino Deep Ganguli ncray86@gmail.com Metamarkets Group, Inc. Metamarkets Group, Inc. gian@metamarkets.com deep@metamarkets.com ABSTRACT event streams into high-value aggregates for a variety of applica- 1 Druid is an open source data store designed for real-time exploratory tions such as business intelligence and A-B testing. analytics on large data sets. The system combines a column-oriented As with many great systems, Hadoop has opened our eyes to storage layout, a distributed, shared-nothing architecture, and an a new space of problems. Specifically, Hadoop excels at storing advanced indexing structure to allow for the arbitrary exploration and providing access to large amounts of data, however, it does not of billion-row tables with sub-second latencies. In this paper, we make any performance guarantees around how quickly that data can describe Druid’s architecture, and detail how it supports fast aggre- be accessed. Furthermore, although Hadoop is a highly available gations, flexible filters, and low latency data ingestion. system, performance degrades under heavy concurrent load. Lastly, while Hadoop works well for storing data, it is not optimized for in- gesting data and making that data immediately readable. Categories and Subject Descriptors Early on in the development of the Metamarkets product, we ran H.2.4 [Database Management]: Systems—Distributed databases into each of these issues and came to the realization that Hadoop is a great back-office, batch processing, and data warehousing sys- tem. However, as a company that has product-level guarantees Keywords around query performance and data availability in a highly concur- distributed; real-time; fault-tolerant; highly available; open source; rent environment (1000+ users), Hadoop wasn’t going to meet our analytics; column-oriented; OLAP needs. We explored different solutions in the space, and after trying both Relational Database Management Systems and NoSQL archi- 1. INTRODUCTION tectures, we came to the conclusion that there was nothing in the open source world that could be fully leveraged for our require- In recent years, the proliferation of internet technology has cre- ments. We ended up creating Druid, an open source, distributed, ated a surge in machine-generated events. Individually, these events column-oriented, real-time analytical data store. In many ways, contain minimal useful information and are of low value. Given the Druid shares similarities with other OLAP systems [30, 35, 22], in- time and resources required to extract meaning from large collec- teractive query systems [28], main-memory databases [14], as well tions of events, many companies were willing to discard this data in- as widely known distributed data stores [7, 12, 23]. The distribution stead. Although infrastructure has been built to handle event-based and query model also borrow ideas from current generation search data (e.g. IBM’s Netezza[37], HP’s Vertica[5], and EMC’s Green- infrastructure [25, 3, 4]. plum[29]), they are largely sold at high price points and are only This paper describes the architecture of Druid, explores the vari- targeted towards those companies who can afford the offering. ous design decisions made in creating an always-on production sys- A few years ago, Google introduced MapReduce [11] as their tem that powers a hosted service, and attempts to help inform any- mechanism of leveraging commodity hardware to index the inter- one who faces a similar problem about a potential method of solving net and analyze logs. The Hadoop [36] project soon followed and it. Druid is deployed in production at several technology compa- was largely patterned after the insights that came out of the original nies2 . The structure of the paper is as follows: we first describe MapReduce paper. Hadoop is currently deployed in many orga- the problem in Section 2. Next, we detail system architecture from nizations to store and analyze large amounts of log data. Hadoop the point of view of how data flows through the system in Section has contributed much to helping companies convert their low-value 3. We then discuss how and why data gets converted into a binary 1 http://druid.io/ https://github.com/metamx/druid format in Section 4. We briefly describe the query API in Section 5 and present performance results in Section 6. Lastly, we leave off Permission to make digital or hard copies of all or part of this work for personal or with our lessons from running Druid in production in Section 7, and classroom use is granted without fee provided that copies are not made or distributed related work in Section 8. for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or 2. PROBLEM DEFINITION republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. Druid was originally designed to solve problems around ingest- SIGMOD’14, June 22–27, 2014, Snowbird, UT, USA. ing and exploring large quantities of transactional events (log data). Copyright is held by the owner/author(s). Publication rights licensed to ACM. This form of timeseries data is commonly found in OLAP work- ACM 978-1-4503-2376-5/14/06 ...$15.00. 2 http://dx.doi.org/10.1145/2588555.2595631. http://druid.io/druid.html

2. Timestamp Page Username Gender City Characters Added Characters Removed 2011-01-01T01:00:00Z Justin Bieber Boxer Male San Francisco 1800 25 2011-01-01T01:00:00Z Justin Bieber Reach Male Waterloo 2912 42 2011-01-01T02:00:00Z Ke$ha Helz Male Calgary 1953 17 2011-01-01T02:00:00Z Ke$ha Xeno Male Taiyuan 3194 170 Table 1: Sample Druid data for edits that have occurred on Wikipedia. flows and the nature of the data tends to be very append heavy. For of each other and there is minimal interaction among them. Hence, example, consider the data shown in Table 1. Table 1 contains data intra-cluster communication failures have minimal impact on data for edits that have occurred on Wikipedia. Each time a user edits availability. a page in Wikipedia, an event is generated that contains metadata To solve complex data analysis problems, the different node types about the edit. This metadata is comprised of 3 distinct compo- come together to form a fully working system. The name Druid nents. First, there is a timestamp column indicating when the edit comes from the Druid class in many role-playing games: it is a was made. Next, there are a set dimension columns indicating var- shape-shifter, capable of taking on many different forms to fulfill ious attributes about the edit such as the page that was edited, the various different roles in a group. The composition of and flow of user who made the edit, and the location of the user. Finally, there data in a Druid cluster are shown in Figure 1. are a set of metric columns that contain values (usually numeric) that can be aggregated, such as the number of characters added or 3.1 Real-time Nodes removed in an edit. Our goal is to rapidly compute drill-downs and aggregates over Real-time nodes encapsulate the functionality to ingest and query this data. We want to answer questions like “How many edits were event streams. Events indexed via these nodes are immediately made on the page Justin Bieber from males in San Francisco?” and available for querying. The nodes are only concerned with events “What is the average number of characters that were added by peo- for some small time range and periodically hand off immutable ple from Calgary over the span of a month?”. We also want queries batches of events they have collected over this small time range to over any arbitrary combination of dimensions to return with sub- other nodes in the Druid cluster that are specialized in dealing with second latencies. batches of immutable events. Real-time nodes leverage Zookeeper The need for Druid was facilitated by the fact that existing open [19] for coordination with the rest of the Druid cluster. The nodes source Relational Database Management Systems (RDBMS) and announce their online state and the data they serve in Zookeeper. NoSQL key/value stores were unable to provide a low latency data Real-time nodes maintain an in-memory index buffer for all in- ingestion and query platform for interactive applications [40]. In the coming events. These indexes are incrementally populated as events early days of Metamarkets, we were focused on building a hosted are ingested and the indexes are also directly queryable. Druid be- dashboard that would allow users to arbitrarily explore and visualize haves as a row store for queries on events that exist in this JVM event streams. The data store powering the dashboard needed to heap-based buffer. To avoid heap overflow problems, real-time return queries fast enough that the data visualizations built on top nodes persist their in-memory indexes to disk either periodically of it could provide users with an interactive experience. or after some maximum row limit is reached. This persist process In addition to the query latency needs, the system had to be multi- converts data stored in the in-memory buffer to a column oriented tenant and highly available. The Metamarkets product is used in a storage format described in Section 4. Each persisted index is im- highly concurrent environment. Downtime is costly and many busi- mutable and real-time nodes load persisted indexes into off-heap nesses cannot afford to wait if a system is unavailable in the face of memory such that they can still be queried. This process is de- software upgrades or network failure. Downtime for startups, who scribed in detail in [33] and is illustrated in Figure 2. often lack proper internal operations management, can determine On a periodic basis, each real-time node will schedule a back- business success or failure. ground task that searches for all locally persisted indexes. The task Finally, another challenge that Metamarkets faced in its early merges these indexes together and builds an immutable block of days was to allow users and alerting systems to be able to make data that contains all the events that have been ingested by a real- business decisions in “real-time”. The time from when an event is time node for some span of time. We refer to this block of data as created to when that event is queryable determines how fast inter- a “segment”. During the handoff stage, a real-time node uploads ested parties are able to react to potentially catastrophic situations in this segment to a permanent backup storage, typically a distributed their systems. Popular open source data warehousing systems such file system such as S3 [12] or HDFS [36], which Druid refers to as as Hadoop were unable to provide the sub-second data ingestion “deep storage”. The ingest, persist, merge, and handoff steps are latencies we required. fluid; there is no data loss during any of the processes. The problems of data exploration, ingestion, and availability span Figure 3 illustrates the operations of a real-time node. The node multiple industries. Since Druid was open sourced in October 2012, starts at 13:37 and will only accept events for the current hour or the it been deployed as a video, network monitoring, operations mon- next hour. When events are ingested, the node announces that it is itoring, and online advertising analytics platform at multiple com- serving a segment of data for an interval from 13:00 to 14:00. Every panies. 10 minutes (the persist period is configurable), the node will flush and persist its in-memory buffer to disk. Near the end of the hour, the node will likely see events for 14:00 to 15:00. When this occurs, 3. ARCHITECTURE the node prepares to serve data for the next hour and creates a new A Druid cluster consists of different types of nodes and each node in-memory index. The node then announces that it is also serving type is designed to perform a specific set of things. We believe a segment from 14:00 to 15:00. The node does not immediately this design separates concerns and simplifies the complexity of the merge persisted indexes from 13:00 to 14:00, instead it waits for overall system. The different node types operate fairly independent a configurable window period for straggling events from 13:00 to

3. Streaming Data Real-time Druid Nodes Nodes External Dependencies MySQL Client Zookeeper Broker Nodes Queries Coordinator Nodes Queries Batch Metadata Deep Historical Data/Segments Data Storage Nodes Figure 1: An overview of a Druid cluster and the flow of data through the cluster. Queries this offset each time they persist their in-memory buffers to disk. In a fail and recover scenario, if a node has not lost disk, it can reload all persisted indexes from disk and continue reading events from Heap and in-memory index Off-heap memory and persisted indexes the last offset it committed. Ingesting events from a recently com- event_34982 event_35789 event_36791 event_23312 event_23481 event_1234 event_2345 event_3456 event_4567 mitted offset greatly reduces a node’s recovery time. In practice, ... event_23593 ... ... ... we see nodes recover from such failure scenarios in a few seconds. event_5678 event_6789 event_7890 event_8901 The second purpose of the message bus is to act as a single end- ... ... Persist point from which multiple real-time nodes can read events. Multi- Disk and persisted indexes ple real-time nodes can ingest the same set of events from the bus, event_1234 event_3456 creating a replication of events. In a scenario where a node com- event_2345 event_4567 Load ... ... pletely fails and loses disk, replicated streams ensure that no data event_5678 event_6789 event_7890 event_8901 is lost. A single ingestion endpoint also allows for data streams ... ... to be partitioned such that multiple real-time nodes each ingest a portion of a stream. This allows additional real-time nodes to be seamlessly added. In practice, this model has allowed one of the largest production Druid clusters to be able to consume raw data at Figure 2: Real-time nodes buffer events to an in-memory index, approximately 500 MB/s (150,000 events/s or 2 TB/hour). which is regularly persisted to disk. On a periodic basis, per- sisted indexes are then merged together before getting handed off. Queries will hit both the in-memory and persisted indexes. 3.2 Historical Nodes Historical nodes encapsulate the functionality to load and serve the immutable blocks of data (segments) created by real-time nodes. In many real-world workflows, most of the data loaded in a Druid 14:00 to arrive. This window period minimizes the risk of data loss cluster is immutable and hence, historical nodes are typically the from delays in event delivery. At the end of the window period, the main workers of a Druid cluster. Historical nodes follow a shared- node merges all persisted indexes from 13:00 to 14:00 into a single nothing architecture and there is no single point of contention among immutable segment and hands the segment off. Once this segment the nodes. The nodes have no knowledge of one another and are is loaded and queryable somewhere else in the Druid cluster, the operationally simple; they only know how to load, drop, and serve real-time node flushes all information about the data it collected for immutable segments. 13:00 to 14:00 and unannounces it is serving this data. Similar to real-time nodes, historical nodes announce their on- line state and the data they are serving in Zookeeper. Instructions to 3.1.1 Availability and Scalability load and drop segments are sent over Zookeeper and contain infor- Real-time nodes are a consumer of data and require a correspond- mation about where the segment is located in deep storage and how ing producer to provide the data stream. Commonly, for data dura- to decompress and process the segment. Before a historical node bility purposes, a message bus such as Kafka [21] sits between the downloads a particular segment from deep storage, it first checks a producer and the real-time node as shown in Figure 4. Real-time local cache that maintains information about what segments already nodes ingest data by reading events from the message bus. The exist on the node. If information about a segment is not present in time from event creation to event consumption is ordinarily on the the cache, the historical node will proceed to download the segment order of hundreds of milliseconds. from deep storage. This process is shown in Figure 5. Once pro- The purpose of the message bus in Figure 4 is two-fold. First, the cessing is complete, the segment is announced in Zookeeper. At message bus acts as a buffer for incoming events. A message bus this point, the segment is queryable. The local cache also allows such as Kafka maintains positional offsets indicating how far a con- for historical nodes to be quickly updated and restarted. On startup, sumer (a real-time node) has read in an event stream. Consumers the node examines its cache and immediately serves whatever data can programmatically update these offsets. Real-time nodes update it finds.

4. 13:47 14:07 ~14:11 persist data for 13:00-14:00 persist data for 13:00-14:00 - unannounce segment for data 13:00-14:00 13:57 persist data for 13:00-14:00 13:00 14:00 15:00 13:37 14:10 - node starts - merge and handoff for data 13:00-14:00 - announce segment - persist data for 14:00-15:00 for data 13:00-14:00 ~14:00 - announce segment for data 14:00-15:00 Figure 3: The node starts, ingests data, persists, and periodically hands data off. This process repeats indefinitely. The time periods between different real-time node operations are configurable. Historical nodes can support read consistency because they only Kafka deal with immutable data. Immutable data blocks also enable a sim- event_12345 ple parallelization model: historical nodes can concurrently scan event_23456 and aggregate immutable blocks without blocking. offset 1 event_34567 event_35582 events Real-time Node 1 3.2.1 Tiers event_37193 Historical nodes can be grouped in different tiers, where all nodes Streaming events event_78901 offset 2 in a given tier are identically configured. Different performance and event_2849219 event_120202 event_79902 fault-tolerance parameters can be set for each tier. The purpose of … event_79932 tiered nodes is to enable higher or lower priority segments to be dis- event_89012 Real-time tributed according to their importance. For example, it is possible event_90192 events events Node 2 to spin up a “hot” tier of historical nodes that have a high num- ber of cores and large memory capacity. The “hot” cluster can be configured to download more frequently accessed data. A parallel Figure 4: Multiple real-time nodes can read from the same mes- “cold” cluster can also be created with much less powerful backing sage bus. Each node maintains its own offset. hardware. The “cold” cluster would only contain less frequently accessed segments. 3.2.2 Availability Historical nodes depend on Zookeeper for segment load and un- Memory load instructions. Should Zookeeper become unavailable, histor- ical nodes are no longer able to serve new data or drop outdated Segment data, however, because the queries are served over HTTP, histori- cal nodes are still able to respond to query requests for the data they are currently serving. This means that Zookeeper outages do not Deep Storage Load impact current data availability on historical nodes. Disk Segment download Segment 3.3 Broker Nodes Broker nodes act as query routers to historical and real-time nodes. create key Broker nodes understand the metadata published in Zookeeper about Cache what segments are queryable and where those segments are located. Entries Broker nodes route incoming queries such that the queries hit the right historical or real-time nodes. Broker nodes also merge partial results from historical and real-time nodes before returning a final consolidated result to the caller. Figure 5: Historical nodes download immutable segments from deep storage. Segments must be loaded in memory before they 3.3.1 Caching can be queried. Broker nodes contain a cache with a LRU [31, 20] invalidation strategy. The cache can use local heap memory or an external dis-

5.tributed key/value store such as Memcached [16]. Each time a bro- and/or a default set of rules may be configured. The coordinator ker node receives a query, it first maps the query to a set of seg- node will cycle through all available segments and match each seg- ments. Results for certain segments may already exist in the cache ment with the first rule that applies to it. and there is no need to recompute them. For any results that do not exist in the cache, the broker node will forward the query to the 3.4.2 Load Balancing correct historical and real-time nodes. Once historical nodes return In a typical production environment, queries often hit dozens or their results, the broker will cache these results on a per segment ba- even hundreds of segments. Since each historical node has limited sis for future use. This process is illustrated in Figure 6. Real-time resources, segments must be distributed among the cluster to en- data is never cached and hence requests for real-time data will al- sure that the cluster load is not too imbalanced. Determining opti- ways be forwarded to real-time nodes. Real-time data is perpetually mal load distribution requires some knowledge about query patterns changing and caching the results is unreliable. and speeds. Typically, queries cover recent segments spanning con- The cache also acts as an additional level of data durability. In tiguous time intervals for a single data source. On average, queries the event that all historical nodes fail, it is still possible to query that access smaller segments are faster. results if those results already exist in the cache. These query patterns suggest replicating recent historical seg- ments at a higher rate, spreading out large segments that are close 3.3.2 Availability in time to different historical nodes, and co-locating segments from In the event of a total Zookeeper outage, data is still queryable. different data sources. To optimally distribute and balance seg- If broker nodes are unable to communicate to Zookeeper, they use ments among the cluster, we developed a cost-based optimization their last known view of the cluster and continue to forward queries procedure that takes into account the segment data source, recency, to real-time and historical nodes. Broker nodes make the assump- and size. The exact details of the algorithm are beyond the scope of tion that the structure of the cluster is the same as it was before the this paper and may be discussed in future literature. outage. In practice, this availability model has allowed our Druid cluster to continue serving queries for a significant period of time 3.4.3 Replication while we diagnosed Zookeeper outages. Coordinator nodes may tell different historical nodes to load a copy of the same segment. The number of replicates in each tier 3.4 Coordinator Nodes of the historical compute cluster is fully configurable. Setups that Druid coordinator nodes are primarily in charge of data manage- require high levels of fault tolerance can be configured to have a ment and distribution on historical nodes. The coordinator nodes high number of replicas. Replicated segments are treated the same tell historical nodes to load new data, drop outdated data, replicate as the originals and follow the same load distribution algorithm. By data, and move data to load balance. Druid uses a multi-version replicating segments, single historical node failures are transparent concurrency control swapping protocol for managing immutable in the Druid cluster. We use this property for software upgrades. segments in order to maintain stable views. If any immutable seg- We can seamlessly take a historical node offline, update it, bring it ment contains data that is wholly obsoleted by newer segments, the back up, and repeat the process for every historical node in a cluster. outdated segment is dropped from the cluster. Coordinator nodes Over the last two years, we have never taken downtime in our Druid undergo a leader-election process that determines a single node that cluster for software upgrades. runs the coordinator functionality. The remaining coordinator nodes act as redundant backups. 3.4.4 Availability A coordinator node runs periodically to determine the current Druid coordinator nodes have Zookeeper and MySQL as external state of the cluster. It makes decisions by comparing the expected dependencies. Coordinator nodes rely on Zookeeper to determine state of the cluster with the actual state of the cluster at the time what historical nodes already exist in the cluster. If Zookeeper be- of the run. As with all Druid nodes, coordinator nodes maintain a comes unavailable, the coordinator will no longer be able to send Zookeeper connection for current cluster information. Coordinator instructions to assign, balance, and drop segments. However, these nodes also maintain a connection to a MySQL database that con- operations do not affect data availability at all. tains additional operational parameters and configurations. One of The design principle for responding to MySQL and Zookeeper the key pieces of information located in the MySQL database is a failures is the same: if an external dependency responsible for co- table that contains a list of all segments that should be served by ordination fails, the cluster maintains the status quo. Druid uses historical nodes. This table can be updated by any service that cre- MySQL to store operational management information and segment ates segments, for example, real-time nodes. The MySQL database metadata information about what segments should exist in the clus- also contains a rule table that governs how segments are created, ter. If MySQL goes down, this information becomes unavailable to destroyed, and replicated in the cluster. coordinator nodes. However, this does not mean data itself is un- available. If coordinator nodes cannot communicate to MySQL, 3.4.1 Rules they will cease to assign new segments and drop outdated ones. Rules govern how historical segments are loaded and dropped Broker, historical, and real-time nodes are still queryable during from the cluster. Rules indicate how segments should be assigned to MySQL outages. different historical node tiers and how many replicates of a segment should exist in each tier. Rules may also indicate when segments should be dropped entirely from the cluster. Rules are usually set 4. STORAGE FORMAT for a period of time. For example, a user may use rules to load the Data tables in Druid (called data sources) are collections of times- most recent one month’s worth of segments into a “hot” cluster, the tamped events and partitioned into a set of segments, where each most recent one year’s worth of segments into a “cold” cluster, and segment is typically 5–10 million rows. Formally, we define a seg- drop any segments that are older. ment as a collection of rows of data that span some period of time. The coordinator nodes load a set of rules from a rule table in the Segments represent the fundamental storage unit in Druid and repli- MySQL database. Rules may be specific to a certain data source cation and distribution are done at a segment level.

6. Cache (on broker nodes) Historical and real-time nodes results for segment 2013-01-01/2013-01-02 segment for data 2013-01-03/2013-01-04 Query for data Query for data from 2013-01-01 results for segment 2013-01-02/2013-01-03 segment for data 2013-01-04/2013-01-05 not in cache to 2013-01-08 results for segment 2013-01-07/2013-01-08 segment for data 2013-01-05/2013-01-06 segment for data 2013-01-06/2013-01-07 Figure 6: Results are cached per segment. Queries combine cached results with results computed on historical and real-time nodes. Druid always requires a timestamp column as a method of sim- plifying data distribution policies, data retention policies, and first- Integer array size (bytes) level query pruning. Druid partitions its data sources into well- Concise compressed size (bytes) defined time intervals, typically an hour or a day, and may further 1e+06 partition on values from other columns to achieve the desired seg- ment size. The time granularity to partition segments is a function sorted of data volume and time range. A data set with timestamps spread sorted unsorted over a year is better partitioned by day, and a data set with times- tamps spread over a day is better partitioned by hour. 1e+04 Segments are uniquely identified by a data source identifer, the time interval of the data, and a version string that increases when- ever a new segment is created. The version string indicates the freshness of segment data; segments with later versions have newer 1e+02 1e+05 1e+08 views of data (over some time range) than segments with older ver- Cardinality sions. This segment metadata is used by the system for concur- rency control; read operations always access data in a particular Figure 7: Integer array size versus Concise set size. time range from the segments with the latest version identifiers for that time range. Druid segments are stored in a column orientation. Given that 4.1 Indices for Filtering Data Druid is best used for aggregating event streams (all data going into In many real world OLAP workflows, queries are issued for the Druid must have a timestamp), the advantages of storing aggregate aggregated results of some set of metrics where some set of di- information as columns rather than rows are well documented [1]. mension specifications are met. An example query is: “How many Column storage allows for more efficient CPU usage as only what Wikipedia edits were done by users in San Francisco who are also is needed is actually loaded and scanned. In a row oriented data male?” This query is filtering the Wikipedia data set in Table 1 based store, all columns associated with a row must be scanned as part of on a Boolean expression of dimension values. In many real world an aggregation. The additional scan time can introduce signficant data sets, dimension columns contain strings and metric columns performance degradations [1]. Druid has multiple column types to represent various data for- contain numeric values. Druid creates additional lookup indices for mats. Depending on the column type, different compression meth- string columns such that only those rows that pertain to a particular ods are used to reduce the cost of storing a column in memory and query filter are ever scanned. on disk. In the example given in Table 1, the page, user, gender, Let us consider the page column in Table 1. For each unique page and city columns only contain strings. Storing strings directly is in Table 1, we can form some representation indicating in which unnecessarily costly and string columns can be dictionary encoded table rows a particular page is seen. We can store this information instead. Dictionary encoding is a common method to compress data in a binary array where the array indices represent our rows. If a and has been used in other data stores such as PowerDrill [17]. In particular page is seen in a certain row, that array index is marked the example in Table 1, we can map each page to a unique integer as 1. For example: identifier. Justin Bieber -> rows [0, 1] -> [1][1][0][0] Justin Bieber -> 0 Ke$ha -> rows [2, 3] -> [0][0][1][1] Ke$ha -> 1 Justin Bieber is seen in rows 0 and 1. This mapping of col- This mapping allows us to represent the page column as an in- umn values to row indices forms an inverted index [39]. To know teger array where the array indices correspond to the rows of the which rows contain Justin Bieber or Ke$ha, we can OR together original data set. For the page column, we can represent the unique the two arrays. pages as follows: [0, 0, 1, 1] [0][1][0][1] OR [1][0][1][0] = [1][1][1][1] The resulting integer array lends itself very well to compression This approach of performing Boolean operations on large bitmap methods. Generic compression algorithms on top of encodings are sets is commonly used in search engines. Bitmap indices for OLAP extremely common in column-stores. Druid uses the LZF [24] com- workloads is described in detail in [32]. Bitmap compression al- pression algorithm. gorithms are a well-defined area of research [2, 44, 42] and often Similar compression methods can be applied to numeric columns. utilize run-length encoding. Druid opted to use the Concise algo- For example, the characters added and characters removed columns rithm [10]. Figure 7 illustrates the number of bytes using Concise in Table 1 can also be expressed as individual arrays. compression versus using an integer array. The results were gen- Characters Added -> [1800, 2912, 1953, 3194] erated on a cc2.8xlarge system with a single thread, 2G heap, Characters Removed -> [25, 42, 17, 170] 512m young gen, and a forced GC between each run. The data set In this case, we compress the raw values as opposed to their dic- is a single day’s worth of data collected from the Twitter garden tionary representations. hose [41] data stream. The data set contains 2,272,295 rows and

7.12 dimensions of varying cardinality. As an additional comparison, [ { we also resorted the data set rows to maximize compression. "timestamp": "2012-01-01T00:00:00.000Z", "result": {"rows":393298} In the unsorted case, the total Concise size was 53,451,144 bytes }, and the total integer array size was 127,248,520 bytes. Overall, { Concise compressed sets are about 42% smaller than integer ar- "timestamp": "2012-01-02T00:00:00.000Z", "result": {"rows":382932} rays. In the sorted case, the total Concise compressed size was }, 43,832,884 bytes and the total integer array size was 127,248,520 ... bytes. What is interesting to note is that after sorting, global com- { "timestamp": "2012-01-07T00:00:00.000Z", pression only increased minimally. "result": {"rows": 1337} } ] 4.2 Storage Engine Druid’s persistence components allows for different storage en- Druid supports many types of aggregations including sums on gines to be plugged in, similar to Dynamo [12]. These storage en- floating-point and integer types, minimums, maximums, and com- gines may store data in an entirely in-memory structure such as the plex aggregations such as cardinality estimation and approximate JVM heap or in memory-mapped structures. The ability to swap quantile estimation. The results of aggregations can be combined storage engines allows for Druid to be configured depending on a in mathematical expressions to form other aggregations. It is be- particular application’s specifications. An in-memory storage en- yond the scope of this paper to fully describe the query API but gine may be operationally more expensive than a memory-mapped more information can be found online3 . storage engine but could be a better alternative if performance is As of this writing, a join query for Druid is not yet implemented. critical. By default, a memory-mapped storage engine is used. This has been a function of engineering resource allocation and use When using a memory-mapped storage engine, Druid relies on case decisions more than a decision driven by technical merit. In- the operating system to page segments in and out of memory. Given deed, Druid’s storage format would allow for the implementation that segments can only be scanned if they are loaded in memory, a of joins (there is no loss of fidelity for columns included as dimen- memory-mapped storage engine allows recent segments to retain sions) and the implementation of them has been a conversation that in memory whereas segments that are never queried are paged out. we have every few months. To date, we have made the choice that The main drawback with using the memory-mapped storage engine the implementation cost is not worth the investment for our organi- is when a query requires more segments to be paged into memory zation. The reasons for this decision are generally two-fold. than a given node has capacity for. In this case, query performance will suffer from the cost of paging segments in and out of memory. 1. Scaling join queries has been, in our professional experience, a constant bottleneck of working with distributed databases. 5. QUERY API Druid has its own query language and accepts queries as POST 2. The incremental gains in functionality are perceived to be requests. Broker, historical, and real-time nodes all share the same of less value than the anticipated problems with managing query API. highly concurrent, join-heavy workloads. The body of the POST request is a JSON object containing key- value pairs specifying various query parameters. A typical query A join query is essentially the merging of two or more streams of will contain the data source name, the granularity of the result data, data based on a shared set of keys. The primary high-level strate- time range of interest, the type of request, and the metrics to ag- gies for join queries we are aware of are a hash-based strategy or a gregate over. The result will also be a JSON object containing the sorted-merge strategy. The hash-based strategy requires that all but aggregated metrics over the time period. one data set be available as something that looks like a hash table, Most query types will also support a filter set. A filter set is a a lookup operation is then performed on this hash table for every Boolean expression of dimension name and value pairs. Any num- row in the “primary” stream. The sorted-merge strategy assumes ber and combination of dimensions and values may be specified. that each stream is sorted by the join key and thus allows for the in- When a filter set is provided, only the subset of the data that per- cremental joining of the streams. Each of these strategies, however, tains to the filter set will be scanned. The ability to handle complex requires the materialization of some number of the streams either in nested filter sets is what enables Druid to drill into data at any depth. sorted order or in a hash table form. The exact query syntax depends on the query type and the infor- When all sides of the join are significantly large tables (> 1 bil- mation requested. A sample count query over a week of data is as lion records), materializing the pre-join streams requires complex follows: distributed memory management. The complexity of the memory { management is only amplified by the fact that we are targeting highly "queryType" : "timeseries", "dataSource" : "wikipedia", concurrent, multitenant workloads. This is, as far as we are aware, "intervals" : "2013-01-01/2013-01-08", an active academic research problem that we would be willing to "filter" : { help resolve in a scalable manner. "type" : "selector", "dimension" : "page", "value" : "Ke$ha" }, 6. PERFORMANCE "granularity" : "day", "aggregations" : [{"type":"count", "name":"rows"}] Druid runs in production at several organizations, and to demon- } strate its performance, we have chosen to share some real world numbers for the main production cluster running at Metamarkets as The query shown above will return a count of the number of rows in the Wikipedia data source from 2013-01-01 to 2013-01-08, fil- of early 2014. For comparison with other databases we also include tered for only those rows where the value of the “page” dimension results from synthetic workloads on TPC-H data. is equal to “Ke$ha”. The results will be bucketed by day and will 3 be a JSON array of the following form: http://druid.io/docs/latest/Querying.html

8. Data Source Dimensions Metrics a 25 21 b 30 26 c 71 35 d 60 19 e 29 8 f 30 16 Mean query latency g 26 18 datasource h a 78 14 1.0 b query time (s) c Table 2: Characteristics of production data sources. d e 0.5 f 6.1 Query Performance in Production g h Druid query performance can vary signficantly depending on the 0.0 query being issued. For example, sorting the values of a high cardi- Feb 03 Feb 10 time Feb 17 Feb 24 nality dimension based on a given metric is much more expensive Query latency percentiles than a simple count over a time range. To showcase the average 1.5 query latencies in a production Druid cluster, we selected 8 of our most queried data sources, described in Table 2. 1.0 90%ile Approximately 30% of queries are standard aggregates involving 0.5 different types of metrics and filters, 60% of queries are ordered datasource group bys over one or more dimensions with aggregates, and 10% 0.0 a 4 of queries are search queries and metadata retrieval queries. The query time (seconds) b number of columns scanned in aggregate queries roughly follows 3 c 95%ile an exponential distribution. Queries involving a single column are 2 d e very frequent, and queries involving all columns are very rare. 1 f A few notes about our results: 0 g • The results are from a “hot” tier in our production cluster. There 20 h were approximately 50 data sources in the tier and several hun- 15 dred users issuing queries. 99%ile 10 • There was approximately 10.5TB of RAM available in the “hot” 5 tier and approximately 10TB of segments loaded. Collectively, 0 Feb 03 Feb 10 Feb 17 Feb 24 there are about 50 billion Druid rows in this tier. Results for every time data source are not shown. Figure 8: Query latencies of production data sources. • The hot tier uses Intel® Xeon® E5-2670 processors and consists of 1302 processing threads and 672 total cores (hyperthreaded). • A memory-mapped storage engine was used (the machine was configured to memory map the data instead of loading it into the Java heap.) Query latencies are shown in Figure 8 and the queries per minute are shown in Figure 9. Across all the various data sources, aver- age query latency is approximately 550 milliseconds, with 90% of queries returning in less than 1 second, 95% in under 2 seconds, and 99% of queries returning in less than 10 seconds. Occasionally Queries per minute we observe spikes in latency, as observed on February 19, where datasource network issues on the Memcached instances were compounded by 1500 a very high query load on one of our largest data sources. b queries / minute c 6.2 Query Benchmarks on TPC-H Data 1000 d We also present Druid benchmarks on TPC-H data. Most TPC-H e queries do not directly apply to Druid, so we selected queries more 500 f typical of Druid’s workload to demonstrate query performance. As g a comparison, we also provide the results of the same queries us- h ing MySQL using the MyISAM engine (InnoDB was slower in our 0 Feb 03 Feb 10 Feb 17 Feb 24 experiments). time We selected MySQL to benchmark against because of its univer- sal popularity. We chose not to select another open source column store because we were not confident we could correctly tune it for Figure 9: Queries per minute of production data sources. optimal performance. Our Druid setup used Amazon EC2 m3.2xlarge instance types (Intel® Xeon® E5-2680 v2 @ 2.80GHz) for historical nodes and c3.2xlarge instances (Intel® Xeon® E5-2670 v2 @ 2.50GHz) for

9. Median query time (100 runs) − 1GB data − single node Druid Scaling − 100GB top_100_parts_filter top_100_parts_details 4 top_100_parts top_100_commitdate Query 3 Time (seconds) sum_price engine Druid sum_all_year 2 MySQL sum_all_filter sum_all count_star_interval 1 0 50 100 150 Time (seconds) 0 Druid Scaling ... 100GB count_star_interval sum_all sum_all_filter sum_all_year sum_price top_100_commitdate top_100_parts top_100_parts_details top_100_parts_filter top_100_parts_filter top_100_parts_details top_100_parts top_100_commitdate Query sum_price Query sum_all_year sum_all_filter sum_all Figure 10: Druid & MySQL benchmarks – 1GB TPC-H data. count_star_interval 1 2 3 4 5 6 Speedup Factor Cores 8 (1 node) 48 (6 nodes) Median Query Time (3+ runs) − 100GB data − single node aggregation top−n 12500 600 10000 Figure 12: Druid scaling benchmarks – 100GB TPC-H data. Time (seconds) 7500 engine 400 Data Source Dimensions Metrics Peak events/s Druid 5000 MySQL s 7 2 28334.60 200 t 10 7 68808.70 2500 u 5 1 49933.93 v 30 10 22240.45 0 0 w 35 14 135763.17 x count_star_interval sum_all sum_all_filter sum_all_year sum_price top_100_commitdate top_100_parts top_100_parts_details top_100_parts_filter 28 6 46525.85 y 33 24 162462.41 z 33 24 95747.74 Query Table 3: Ingestion characteristics of various data sources. Figure 11: Druid & MySQL benchmarks – 100GB TPC-H data. Note that in this setup, several other data sources were being in- broker nodes. Our MySQL setup was an Amazon RDS instance gested and many other Druid related ingestion tasks were running that ran on the same m3.2xlarge instance type. concurrently on the machines. The results for the 1 GB TPC-H data set are shown in Figure 10 Druid’s data ingestion latency is heavily dependent on the com- and the results of the 100 GB data set are shown in Figure 11. plexity of the data set being ingested. The data complexity is de- We benchmarked Druid’s scan rate at 53,539,211 rows/second/core termined by the number of dimensions in each event, the number for select count(*) equivalent query over a given time interval of metrics in each event, and the types of aggregations we want to and 36,246,530 rows/second/core for a select sum(float) type perform on those metrics. With the most basic data set (one that query. only has a timestamp column), our setup can ingest data at a rate of Finally, we present our results of scaling Druid to meet increasing 800,000 events/second/core, which is really just a measurement of data volumes with the TPC-H 100 GB data set. We observe that how fast we can deserialize events. Real world data sets are never when we increased the number of cores from 8 to 48, not all types of this simple. Table 3 shows a selection of data sources and their queries achieve linear scaling, but the simpler aggregation queries characteristics. do, as shown in Figure 12. We can see that, based on the descriptions in Table 3, latencies The increase in speed of a parallel computing system is often lim- vary significantly and the ingestion latency is not always a factor of ited by the time needed for the sequential operations of the system. the number of dimensions and metrics. We see some lower latencies In this case, queries requiring a substantial amount of work at the on simple data sets because that was the rate that the data producer broker level do not parallelize as well. was delivering data. The results are shown in Figure 13. We define throughput as the number of events a real-time node can ingest and also make queryable. If too many events are sent 6.3 Data Ingestion Performance to the real-time node, those events are blocked until the real-time To showcase Druid’s data ingestion latency, we selected several node has capacity to accept them. The peak ingestion latency we production datasources of varying dimensions, metrics, and event measured in production was 22914.43 events/second/core on a data- volumes. Our production ingestion setup consists of 6 nodes, to- source with 30 dimensions and 19 metrics, running an Amazon talling 360GB of RAM and 96 cores (12 x Intel® Xeon® E5-2670). cc2.8xlarge instance.

10. Events per second ... 24h moving average Data Center Outages. 250,000 datasource Complete cluster failures are possible, but extremely rare. If s Druid is only deployed in a single data center, it is possible for 200,000 t the entire data center to fail. In such cases, new machines need 150,000 u to be provisioned. As long as deep storage is still available, clus- events / s v ter recovery time is network bound, as historical nodes simply need 100,000 w to redownload every segment from deep storage. We have experi- x enced such failures in the past, and the recovery time was several 50,000 y hours in the Amazon AWS ecosystem for several terabytes of data. 0 z 7.1 Operational Monitoring Dec 15 Jan 01 Jan 15 time Feb 01 Feb 15 Mar 01 Proper monitoring is critical to run a large scale distributed clus- ter. Each Druid node is designed to periodically emit a set of oper- ational metrics. These metrics may include system level data such Figure 13: Combined cluster ingestion rates. as CPU usage, available memory, and disk capacity, JVM statistics such as garbage collection time, and heap usage, or node specific metrics such as segment scan time, cache hit rates, and data inges- The latency measurements we presented are sufficient to address tion latencies. Druid also emits per query metrics. the stated problems of interactivity. We would prefer the variability We emit metrics from a production Druid cluster and load them in the latencies to be less. It is still possible to decrease latencies into a dedicated metrics Druid cluster. The metrics Druid cluster by adding additional hardware, but we have not chosen to do so is used to explore the performance and stability of the production because infrastructure costs are still a consideration for us. cluster. This dedicated metrics cluster has allowed us to find nu- merous production problems, such as gradual query speed degrega- tions, less than optimally tuned hardware, and various other system bottlenecks. We also use a metrics cluster to analyze what queries 7. DRUID IN PRODUCTION are made in production and what aspects of the data users are most Over the last few years, we have gained tremendous knowledge interested in. about handling production workloads with Druid and have made a couple of interesting observations. 7.2 Pairing Druid with a Stream Processor Currently, Druid can only understand fully denormalized data Query Patterns. streams. In order to provide full business logic in production, Druid Druid is often used to explore data and generate reports on data. can be paired with a stream processor such as Apache Storm [27]. In the explore use case, the number of queries issued by a single A Storm topology consumes events from a data stream, retains user are much higher than in the reporting use case. Exploratory only those that are “on-time”, and applies any relevant business queries often involve progressively adding filters for the same time logic. This could range from simple transformations, such as id range to narrow down results. Users tend to explore short time in- to name lookups, to complex operations such as multi-stream joins. tervals of recent data. In the generate report use case, users query The Storm topology forwards the processed event stream to Druid for much longer data intervals, but those queries are generally few in real-time. Storm handles the streaming data processing work, and pre-determined. and Druid is used for responding to queries for both real-time and historical data. Multitenancy. 7.3 Multiple Data Center Distribution Expensive concurrent queries can be problematic in a multitenant environment. Queries for large data sources may end up hitting ev- Large scale production outages may not only affect single nodes, ery historical node in a cluster and consume all cluster resources. but entire data centers as well. The tier configuration in Druid co- Smaller, cheaper queries may be blocked from executing in such ordinator nodes allow for segments to be replicated across multiple cases. We introduced query prioritization to address these issues. tiers. Hence, segments can be exactly replicated across historical Each historical node is able to prioritize which segments it needs nodes in multiple data centers. Similarily, query preference can be to scan. Proper query planning is critical for production workloads. assigned to different tiers. It is possible to have nodes in one data Thankfully, queries for a significant amount of data tend to be for center act as a primary cluster (and receive all queries) and have a reporting use cases and can be deprioritized. Users do not expect redundant cluster in another data center. Such a setup may be de- the same level of interactivity in this use case as when they are ex- sired if one data center is situated much closer to users. ploring data. 8. RELATED WORK Node failures. Cattell [6] maintains a great summary about existing Scalable Single node failures are common in distributed environments, but SQL and NoSQL data stores. Hu [18] contributed another great many nodes failing at once are not. If historical nodes completely summary for streaming databases. Druid feature-wise sits some- fail and do not recover, their segments need to be reassigned, which where between Google’s Dremel [28] and PowerDrill [17]. Druid means we need excess cluster capacity to load this data. The amount has most of the features implemented in Dremel (Dremel handles of additional capacity to have at any time contributes to the cost arbitrary nested data structures while Druid only allows for a single of running a cluster. From our experiences, it is extremely rare to level of array-based nesting) and many of the interesting compres- see more than 2 nodes completely fail at once and hence, we leave sion algorithms mentioned in PowerDrill. enough capacity in our cluster to completely reassign the data from Although Druid builds on many of the same principles as other 2 historical nodes. distributed columnar data stores [15], many of these data stores are

11.designed to be more generic key-value stores [23] and do not sup- [5] C. Bear, A. Lamb, and N. Tran. The vertica database: Sql port computation directly in the storage layer. There are also other rdbms for managing big data. In Proceedings of the 2012 data stores designed for some of the same data warehousing issues workshop on Management of big data systems, pages 37–38. that Druid is meant to solve. These systems include in-memory ACM, 2012. databases such as SAP’s HANA [14] and VoltDB [43]. These data [6] R. Cattell. Scalable sql and nosql data stores. ACM SIGMOD stores lack Druid’s low latency ingestion characteristics. Druid also Record, 39(4):12–27, 2011. has native analytical features baked in, similar to ParAccel [34], [7] F. Chang, J. Dean, S. Ghemawat, W. C. Hsieh, D. A. however, Druid allows system wide rolling software updates with Wallach, M. Burrows, T. Chandra, A. Fikes, and R. E. no downtime. Gruber. Bigtable: A distributed storage system for structured Druid is similiar to C-Store [38] and LazyBase [8] in that it has data. ACM Transactions on Computer Systems (TOCS), two subsystems, a read-optimized subsystem in the historical nodes 26(2):4, 2008. and a write-optimized subsystem in real-time nodes. Real-time nodes [8] J. Cipar, G. Ganger, K. Keeton, C. B. Morrey III, C. A. are designed to ingest a high volume of append heavy data, and do Soules, and A. Veitch. Lazybase: trading freshness for not support data updates. Unlike the two aforementioned systems, performance in a scalable database. In Proceedings of the 7th Druid is meant for OLAP transactions and not OLTP transactions. ACM european conference on Computer Systems, pages Druid’s low latency data ingestion features share some similar- 169–182. ACM, 2012. ities with Trident/Storm [27] and Spark Streaming [45], however, [9] Cloudera impala. http://blog.cloudera.com/blog, both systems are focused on stream processing whereas Druid is March 2013. focused on ingestion and aggregation. Stream processors are great [10] A. Colantonio and R. Di Pietro. Concise: Compressed complements to Druid as a means of pre-processing the data before ‘n’composable integer set. Information Processing Letters, the data enters Druid. 110(16):644–650, 2010. There are a class of systems that specialize in queries on top of cluster computing frameworks. Shark [13] is such a system for [11] J. Dean and S. Ghemawat. Mapreduce: simplified data processing on large clusters. Communications of the ACM, queries on top of Spark, and Cloudera’s Impala [9] is another system focused on optimizing query performance on top of HDFS. Druid 51(1):107–113, 2008. historical nodes download data locally and only work with native [12] G. DeCandia, D. Hastorun, M. Jampani, G. Kakulapati, Druid indexes. We believe this setup allows for faster query laten- A. Lakshman, A. Pilchin, S. Sivasubramanian, P. Vosshall, cies. and W. Vogels. Dynamo: amazon’s highly available Druid leverages a unique combination of algorithms in its archi- key-value store. In ACM SIGOPS Operating Systems tecture. Although we believe no other data store has the same set Review, volume 41, pages 205–220. ACM, 2007. of functionality as Druid, some of Druid’s optimization techniques [13] C. Engle, A. Lupher, R. Xin, M. Zaharia, M. J. Franklin, such as using inverted indices to perform fast filters are also used in S. Shenker, and I. Stoica. Shark: fast data analysis using other data stores [26]. coarse-grained distributed memory. In Proceedings of the 2012 international conference on Management of Data, pages 689–692. ACM, 2012. 9. CONCLUSIONS [14] F. Färber, S. K. Cha, J. Primsch, C. Bornhövd, S. Sigg, and In this paper we presented Druid, a distributed, column-oriented, W. Lehner. Sap hana database: data management for modern real-time analytical data store. Druid is designed to power high business applications. ACM Sigmod Record, 40(4):45–51, performance applications and is optimized for low query latencies. 2012. Druid supports streaming data ingestion and is fault-tolerant. We [15] B. Fink. Distributed computation on dynamo-style discussed Druid benchmarks and summarized key architecture as- distributed storage: riak pipe. In Proceedings of the eleventh pects such as the storage format, query language, and general exe- ACM SIGPLAN workshop on Erlang workshop, pages cution. 43–50. ACM, 2012. [16] B. Fitzpatrick. Distributed caching with memcached. Linux 10. ACKNOWLEDGEMENTS journal, (124):72–74, 2004. Druid could not have been built without the help of many great [17] A. Hall, O. Bachmann, R. Büssow, S. Gănceanu, and engineers at Metamarkets and in the community. We want to thank M. Nunkesser. Processing a trillion cells per mouse click. everyone that has contributed to the Druid codebase for their in- Proceedings of the VLDB Endowment, 5(11):1436–1446, valuable support. 2012. [18] B. Hu. Stream database survey. 2011. 11. REFERENCES [19] P. Hunt, M. Konar, F. P. Junqueira, and B. Reed. Zookeeper: [1] D. J. Abadi, S. R. Madden, and N. Hachem. Column-stores Wait-free coordination for internet-scale systems. In USENIX vs. row-stores: How different are they really? In Proceedings ATC, volume 10, 2010. of the 2008 ACM SIGMOD international conference on [20] C. S. Kim. Lrfu: A spectrum of policies that subsumes the Management of data, pages 967–980. ACM, 2008. least recently used and least frequently used policies. IEEE [2] G. Antoshenkov. Byte-aligned bitmap compression. In Data Transactions on Computers, 50(12), 2001. Compression Conference, 1995. DCC’95. Proceedings, page [21] J. Kreps, N. Narkhede, and J. Rao. Kafka: A distributed 476. IEEE, 1995. messaging system for log processing. In Proceedings of 6th [3] Apache. Apache solr. International Workshop on Networking Meets Databases http://lucene.apache.org/solr/, February 2013. (NetDB), Athens, Greece, 2011. [4] S. Banon. Elasticsearch. [22] T. Lachev. Applied Microsoft Analysis Services 2005: And http://www.elasticseach.com/, July 2013. Microsoft Business Intelligence Platform. Prologika Press, 2005.

12.[23] A. Lakshman and P. Malik. Cassandra—a decentralized [36] K. Shvachko, H. Kuang, S. Radia, and R. Chansler. The structured storage system. Operating systems review, hadoop distributed file system. In Mass Storage Systems and 44(2):35, 2010. Technologies (MSST), 2010 IEEE 26th Symposium on, pages [24] Liblzf. http://freecode.com/projects/liblzf, March 1–10. IEEE, 2010. 2013. [37] M. Singh and B. Leonhardi. Introduction to the ibm netezza [25] LinkedIn. Senseidb. http://www.senseidb.com/, July warehouse appliance. In Proceedings of the 2011 Conference 2013. of the Center for Advanced Studies on Collaborative [26] R. MacNicol and B. French. Sybase iq multiplex-designed Research, pages 385–386. IBM Corp., 2011. for analytics. In Proceedings of the Thirtieth international [38] M. Stonebraker, D. J. Abadi, A. Batkin, X. Chen, conference on Very large data bases-Volume 30, pages M. Cherniack, M. Ferreira, E. Lau, A. Lin, S. Madden, 1227–1230. VLDB Endowment, 2004. E. O’Neil, et al. C-store: a column-oriented dbms. In [27] N. Marz. Storm: Distributed and fault-tolerant realtime Proceedings of the 31st international conference on Very computation. http://storm-project.net/, February large data bases, pages 553–564. VLDB Endowment, 2005. 2013. [39] A. Tomasic and H. Garcia-Molina. Performance of inverted [28] S. Melnik, A. Gubarev, J. J. Long, G. Romer, S. Shivakumar, indices in shared-nothing distributed text document M. Tolton, and T. Vassilakis. Dremel: interactive analysis of information retrieval systems. In Parallel and Distributed web-scale datasets. Proceedings of the VLDB Endowment, Information Systems, 1993., Proceedings of the Second 3(1-2):330–339, 2010. International Conference on, pages 8–17. IEEE, 1993. [29] D. Miner. Unified analytics platform for big data. In [40] E. Tschetter. Introducing druid: Real-time analytics at a Proceedings of the WICSA/ECSA 2012 Companion Volume, billion rows per second. http://druid.io/blog/2011/ pages 176–176. ACM, 2012. 04/30/introducing-druid.html, April 2011. [30] K. Oehler, J. Gruenes, C. Ilacqua, and M. Perez. IBM Cognos [41] Twitter public streams. https://dev.twitter.com/ TM1: The Official Guide. McGraw-Hill, 2012. docs/streaming-apis/streams/public, March 2013. [31] E. J. O’neil, P. E. O’neil, and G. Weikum. The lru-k page [42] S. J. van Schaik and O. de Moor. A memory efficient replacement algorithm for database disk buffering. In ACM reachability data structure through bit vector compression. In SIGMOD Record, volume 22, pages 297–306. ACM, 1993. Proceedings of the 2011 international conference on Management of data, pages 913–924. ACM, 2011. [32] P. O’Neil and D. Quass. Improved query performance with variant indexes. In ACM Sigmod Record, volume 26, pages [43] L. VoltDB. Voltdb technical overview. 38–49. ACM, 1997. https://voltdb.com/, 2010. [33] P. O’Neil, E. Cheng, D. Gawlick, and E. O’Neil. The [44] K. Wu, E. J. Otoo, and A. Shoshani. Optimizing bitmap log-structured merge-tree (lsm-tree). Acta Informatica, indices with efficient compression. ACM Transactions on 33(4):351–385, 1996. Database Systems (TODS), 31(1):1–38, 2006. [34] Paraccel analytic database. [45] M. Zaharia, T. Das, H. Li, S. Shenker, and I. Stoica. http://www.paraccel.com/resources/Datasheets/ Discretized streams: an efficient and fault-tolerant model for ParAccel-Core-Analytic-Database.pdf, March 2013. stream processing on large clusters. In Proceedings of the 4th USENIX conference on Hot Topics in Cloud Computing, [35] M. Schrader, D. Vlamis, M. Nader, C. Claterbos, D. Collins, pages 10–10. USENIX Association, 2012. M. Campbell, and F. Conrad. Oracle Essbase & Oracle OLAP. McGraw-Hill, Inc., 2009.