WO2014107359A1

WO2014107359A1 - System and method for distributed database query engines

Info

Publication number: WO2014107359A1
Application number: PCT/US2013/077240
Authority: WO
Inventors: Raghotham Murthy; Rajat Goel
Original assignee: Facebook Inc
Current assignee: Meta Platforms Inc
Priority date: 2013-01-07
Filing date: 2013-12-20
Publication date: 2014-07-10
Anticipated expiration: 2015-07-07
Also published as: CA2896855A1; US20190121809A1; MX349042B; EP3096250B1; US10698913B2; JP2016509294A; EP2752779B1; CN104903894A; CN104903894B; US20140195558A1; JP6117378B2; KR20150104585A; US11347761B1; US20160188677A1; EP3096250A1; US9361344B2; AU2013371448B2; CN110069526A; US9081826B2; KR101775433B1

Abstract

Techniques for a system capable of performing low latency database query processing are disclosed herein. The system includes a gateway server and a plurality of worker nodes. The gateway server is configured to divide a database query, for a database containing data stored in a distributed storage cluster having a plurality of data nodes, into a plurality of partial queries and construct a query result based on a plurality of intermediate results. Each worker node of the plurality of worker nodes is configured to process a respective partial query of the plurality of partial queries by scanning data related to the respective partial query that stored on at least one data node of the distributed storage cluster and generate an intermediate result of the plurality of intermediate results that is stored in a memory of that worker node.

Description

SYSTEM AN D METHOD FOR DISTRIBUTED DATABASE

QU ERY ENGI NES

CROSS-REFERENCE TO RELATED APPLICATION

[0001 ] This application claims priority to U.S. Patent Application No. 13/735,820 filed 7 January 2013, which is hereby incorporated by reference in its entirety.

[0002] This application claims priority to European Patent Application No. 13198563.2 filed 19 December 2013, which is hereby incorporated by reference in its entirety. FIE LD OF THE INVENTION

[0003] This invention relates generally to databases, and in particular to a distributed database query engine for low query latency database analyses. BACKGROU N D

[0004] Developments in computer and networking technology have given rise to applications that requ ire massive amounts of data storage. For example, tens of millions of users can create web pages and u pload images and text to a social media website. Consequently, a social media website can accumulate massive amounts of data each day and therefore need a highly scalable system for storing and processing data. Various tools exist to facil itate such mass data storage.

Frameworks exist that support large-scale data-intensive distributed applications, by enabling applications to interact with a cluster of

thousands of computers (also referred to as nodes) and petabytes of data. For instance, a framework called Hadoop util izes a distributed , scalable, portable fi le system , called Hadoop Distributed File System (H DFS), to distribute a massive amount of data among data nodes (also referred to as slave nodes) in a Hadoop cluster. I n order to reduce the adverse impact of a data node power outage or network failure (including switch failu re), data in an H DFS is typically replicated on different data nodes. [0005] Hive, an open source data warehouse system , was developed to run on top of Hadoop clusters. Hive su pports data queries expressed in a scripted query language (SQL)-like declarative language called HiveQL. The H ive system then compiles the queries expressed in HiveQL into map-reduce jobs that can be executed on the Hadoop cluster, in a mathematical form of directed acyclic graph . The HiveQL language includes a type system that supports tables containing primitive types, col lections such as arrays and maps, and nested compositions of types. I n addition , the Hive system includes a system catalog , called

H ive-Metastore, containing schemes and statistics, which is useful in data exploration and query optimization .

[0006] Coupled with the Hadoop cluster, the H ive system can store and analyze large amounts of data for a social networking system . For example, the H ive system can analyze the degree of connection between users to rank stories that users follow on the social networking system . The H ive system can analyze activity logs to gain insights into how services of the social networking system are bei ng used to help application developers, page administrators and advertisers make development and business decisions. The H ive system can run complex data mining programs to optim ize the advertisements shown to the users of the social networking system . The H ive system can further analyze the usage logs to identify spam and abuse of the social networki ng system .

[0007] The H ive system incl udes web-based tools for people without programm ing ability to author and execute H ive queries, for authori ng , debugging and scheduling complex data pipelines, and for generating reports based on data stored in the H ive system and other relational databases like MySQL and Oracle.

[0008] However, query latency for the H ive system is usually h igh . Due to the large amou nt of data and the map-reduce scheme of the Hadoop cluster, even a simplest query can take from several seconds to minutes to complete. This is particular a problem for interactive analyses when an operator needs the result of the current query to decide the next query of a series of queries. The latency problem significantly affects the productivity of the analysts since the analysts cannot determine the next query when waiting for the result of the current query.

[0009] One possible workaround solution is to create data pipelines that load aggregate data from Hive into other type of relational database management system (RDB MS) such as MySQL and Oracle. Then the operator performs interactive analysis and bui lds reports using these RDB MS . However, each RDB MS needs a separate data pipeline. It also takes time for the data pipeline to transfer the aggregate data from H ive to other RDBMS. Thus, this workaround process is sti ll cumbersome and inconvenient.

SU MMARY

[001 0] The technology introduced here provides the benefits of low-latency queries on a very large amount of data stored in a large scale storage cluster (e.g. a Hadoop cluster), which stores metadata in a system catalog (e.g . Hive Metastore). I n particular, the technology introduced here incl udes a distributed query engi ne based on a servi ng tree

computation framework. The distributed query engine includes a gateway server and a plural ity of work nodes. The gateway server divides a query task into partial tasks. Each worker node of the engine handles a partial task to generate an intermediate query result in memory. The intermediate query result can be an approximate intermediate result generated by scanning a portion of the related data. The gateway server receives the intermediate query results and constructs a query result for the query task based on the intermediate query results.

[001 1 ] I n accordance with the techniques i ntroduced here, therefore, a system for processing a database query for a database is provided . The system includes a gateway server and a plu rality of worker nodes. The gateway server is configured to divide the database query, for the database containing data stored in a distributed storage cluster havi ng a plurality of data nodes, into a plurality of partial queries and construct a query result based on a plurality of intermediate results. Each worker node of the plurality of worker nodes is configured to process a respective partial query of the plurality of partial queries by scanning data related to the respective partial query that stored on at least one data node of the distributed storage cluster and generate an intermediate result of the plurality of i ntermediate results that is stored in a memory of that worker node.

[0012] The techniques introduced here are capable of performing low-latency query processing on a large aggregate of data stored in a large scale storage cluster such as a Hadoop cluster. Th is is particularly beneficial for interactive analyses since operators do not need to wait for the cu rrent query to complete to determine the next query. This distributed query system can further generate approximate results by scanning a portion of the related data. Operators of the system can receive rapid prototyping of a series of queries to test the valid ity of the series the queries, before processing the series of queries on the whole set of the related data.

[001 3] Embodiments according to the invention are in particular disclosed in the attached claims di rected to a system, a storage medium and a method, wherein any feature mentioned in one claim category, e .g. system , can be claimed in another claim category, e.g . method , as well .

[0014] I n an embodiment of the invention , a system comprises:

a gateway server configured to generate a pl urality of partial queries from a database query for a database containing data stored in a distributed storage cluster that has a plurality of data nodes, and to construct a query result based on a plural ity of intermediate results; and a plural ity of worker nodes, wherein each worker node of the plurality of worker nodes is configured to process a respective partial query of the plurality of partial queries by scanning data related to the respective partial query and stored on at least one data node of the distributed storage cluster, and wherein each worker node of the pl urality of worker nodes is further configured to generate an intermediate resu lt of the plu rality of intermediate results that is stored in a memory of that worker node. [001 5] Each worker node of the plurality of worker nodes can be further configured to process the respective partial query of the plu rality of partial queries by scanni ng a portion of the data related to the respective partial query that is stored on the at least one data node of the distributed storage cluster and to generate an approximate intermediate result that is stored in the memory of that worker node.

[001 6] The gateway server can be further configured to construct an approximate query result based on at least one approximate intermed iate result.

[001 7] The gateway server can also further be configured to construct an approximate query result based on a portion of the plurality of intermediate results.

[001 8] The gateway server can even further be configured to identify a straggling worker node, further d ivide a partial query that is assigned to the straggl ing worker node i nto a plu rality of subordinate partial queries, and assign the plurality of subordinate partial queries to some of the plurality of worker nodes, wherein the straggling worker node is a worker node that either fails to report a rate of progress to the gateway server or reports the rate of progress below a predetermined val ue after a

predetermined time period to the gateway server.

[001 9] Each worker node of the plurality of the worker nodes can be a service running a respective data node within the distributed storage cluster.

[0020] The system further can com prise:

a metadata cache configured to cache table level metadata of the database and fi le level metadata of the distributed storage cluster.

[0021 ] The metadata cache can be configured to retain cached metadata from a previous database query for the database query.

[0022] Each worker node of the plurality of the worker nodes can periodical ly send heartbeat messages to the gateway server to report status of a partial query processing by that worker node. [0023] The gateway server can further be configured to receive an instruction from a client device to retu rn an approximate query result or terminate a processing of the database query.

[0024] The gateway server can also further be configured to instruct the worker nodes to immed iately return approxi mate intermediate results, and to return an approximate query result based on the approximate

intermediate results to a client device.

[0025] The database query can include a request for an approximate query result.

[0026] The query result can be accompanied by an indication of a portion of related data stored in the data nodes that has been scanned for the query result.

[0027] The database can be a Hive data warehouse system and the distributed storage cluster is a Hadoop cl uster.

[0028] I n a further embodiment of the invention , a method uses a system according to the invention or any of the above mentioned em bodiments.

[0029] I n a further embodiment of the invention a method comprises:

receiving from a client device a database query for a database containing data stored in a distributed storage cluster that has a plurality of data nodes;

dividing the database query into a plurality of partial queries; sending each of the partial queries to a respective worker node of a plural ity of worker nodes, wherein each worker node is a service runn ing on a data node of the distributed storage cluster;

retrieving a plu rality of intermediate resu lts for the partial queries from the worker nodes, wherein each intermediate resu lt is processed by a respective worker node of the worker nodes by scanning related data stored in a data node on which the perspective worker node ru ns; and

generating a query result based on the plurality of intermediate results.

[0030] The method further can comprise: returning the query result along with a portion indicator to the client device, wherein the portion indicator indicates the portion of related data stored in the data nodes that has been scanned for the query result.

[0031 ] The method also further can comprise :

instructing the worker nodes to im mediately return approximate query results;

and wherein the step of retrieving comprises:

retrieving a plurality of approximate intermediate results for the partial queries from the worker nodes, wherein each approxi mate

intermediate result is processed by a respective worker node of the worker nodes by scanning a portion of related data stored in a data node on which the perspective worker node runs.

[0032] The method may further comprise:

for each partial query, retrievi ng metadata regarding which data node stores data related to the partial query;

and the step of sending comprises:

sending each of the partial queries to a respective worker node of a plural ity of worker nodes based on the metadata.

[0033] I n a further embodiment of the invention , which can be claimed as well , a method comprises:

receiving a database query from a client device , for a database containing data stored in a distributed storage cluster havi ng a plural ity of data nodes;

identifying a straggling worker node, dividing a partial query that is assigned to the straggling worker node into a plurality of su bordinate partial queries, and assigning the plural ity of subord inate partial queries to some of the plurality of worker nodes;

retrieving a plurality of intermediate resu lts for the partial queries from the worker nodes, wherein each intermediate result is processed by a respective worker node of the worker nodes by scanning related data stored in a data node on which the perspective worker node ru ns; and

generating a query result based on the plurality of intermediate results.

[0034] The step of identifying can comprise :

identifying a straggling worker node by monitoring heartbeat messages that the worker nodes periodically send , wherein the straggli ng worker node is identified when heartbeat messages from the straggling worker node are not received for a predetermined ti me period , or when a heartbeat message from the straggling worker node is received and the heartbeat message includ ing a number representing a status of a partial query processing by the straggli ng worker node that is below a threshold value.

[0035] I n a further embodiment of the invention , which can be claimed as well , one or more computer-readable non-transitory storage media embody software that is operable when executed to perform in a system according to the invention or any of the above mentioned em bodiments.

[0036] Other aspects of the technology introduced here will be apparent from the accompanyi ng figures and from the detailed description, which follows.

BRIEF DESCRIPTION OF THE DRAWI NGS

[0037] These and other objects, features and characteristics of the present invention wil l become more apparent to those ski lled in the art from a study of the following detailed description in conjunction with the appended claims and drawings, all of wh ich form a part of this

specification . I n the drawings:

FIG . 1 illustrates an example of a Hadoop cluster upon which a distributed query engine can be built.

FIG . 2 illustrates an example of a Hadoop cluster having a

JobTracker managing a MapReduce task. FIG . 3 ill ustrates relationships among a distributed query engine, a Hadoop Distributed File System (H DFS), a H ive data warehouse and a storage cluster.

FIG . 4 illustrates a h igh level block diagram of an example distributed query engine.

FIG . 5 ill ustrates a sample process for identifyi ng straggling worker nodes and further dividing the partial queries.

FIG . 6 ill ustrates a sample process for approximate processing of a database query.

F IG . 7 is a high-level block diagram showi ng an example of the architecture of a computer node, wh ich may represent any cluster node described herein .

DETAI LED DESCRIPTION

[0038] References in th is description to "an embodiment", "one embodiment", or the like, mean that the particular feature, function, or characteristic being described is included in at least one embodiment of the present invention . Occurrences of such phrases in this description do not necessarily al l refer to the same embodiment, nor are they necessari ly mutual ly exclusive.

[0039] Modern social networking systems can accumulate massive amou nts of data each day and therefore need a highly scalable system for storing and analyzi ng data. Particularly, efficient interactive analyses on the massive amount of data need a low-latency fast-responding way of processing data queries. The present invention discloses a distributed query engine enabled by com bin ing an in-memory serving-tree based computation framework with approximate query processing . The

distributed query engine d ivides a query task into m ultiple partial tasks and distributes the partial tasks to worker nodes for further i n-memory processing. The distributed query engi ne is capable of generating an approximate result at any time during the query processing , by requesting approximate intermediate results from the worker nodes based on the scanned portion of the data. U nlike the map-reduce scheme of the conventional Hadoop clusters, the worker nodes process the partial tasks and store the entire intermediate results in memory to reduce the

processing time and improve overal l latency. Only the intermed iate results, instead of the underlying date , are transferred for the construction of the result, significantly reducing the amount of transferred data and the transferring time.

[0040] I n one embodiment, the d istributed query engine can be built on top of a Hadoop cluster runn ing Hadoop Distributed File System (H DFS), H ive data warehouse, and H ive-Metastore. The distributed query engine can be com patible with H ive's data formats and metadata, and support a subset of HiveQL language. Operators using the distributed query engine can efficiently discover statistics patterns in the data managed by the H ive data warehouse. The distri buted query engine can conduct quick analyses and rapid prototyping of a series of queries by generating approximate results. In addition , the distributed query engine can run fu ll analyses by scanning the entire related data aggregate.

[0041 ] FIG . 1 illustrates an example of a Hadoop cluster that a

distributed query engine can be built upon . I n Figure 1 , the Hadoop cluster 1 00 includes a metadata node 1 1 0A and a plu rality of data nodes 1 1 0B, 1 1 0C and 1 1 0D. The nodes can communicate with each other through an interconnect 1 20. The interconnect 1 20 may be , for example, a local area network (LAN ), wide area network (WAN), metropolitan area network (MAN ), global area network such as the I nternet, a Fibre Channel fabric, or any combination of such interconnects. I n some em bodiments, the interconnect 1 20 can include a network switch for processing and routing data between the nodes under network protocols, incl udi ng TCP/I P.

Clients 1 30A and 1 30 B may comm unicate with the Hadoop cluster 1 00 via a network 1 40, which can be, for exam ple, the I nternet, a LAN , or any other type of network or combination of networks. Each of the clients may be, for example, a conventional personal com puter (PC), server-class computer, workstation , handheld computing/communication device, or the like. In some embodiments, the Hadoop cluster is implemented using one or more racks of commod ity-class servers. [0042] Files and data are stored in a distributed manner across the nodes in the Hadoop cluster 100 in a Hadoop Distributed File System (HDFS). To the clients of the cluster 100, including clients 130A and 130B, the HDFS serves the functions of a conventional hierarchical file system. Files and data blocks of files can be created, deleted, or moved in the HDFS. There is a NameNode service 150 running on the metadata node 110A to provide metadata services within HDFS, including

maintaining the file system namespace and controlling access by external clients. The NameNode service can store the file system index, including the mapping of blocks to files and file system properties, in a file 160 called Fslmage in the metadata node. In some embodiments, there may be a secondary metadata node running a secondary NameNode service. The secondary metadata node serves as a backup in case the metadata node fails.

[0043] Each data node 110 is responsible for storing files for the HDFS. Files stored in the HDFS are divided into subsets, referred to herein as "blocks". In one embodiment, the size of a block is 64 MBs. The blocks are typically replicated to multiple data nodes. Therefore, the HDFS in the Hadoop cluster 100 can, but does not necessarily, use a conventional RAID architecture to achieve data reliability. File operations are controlled by the NameNode service 150 running on the metadata node 110A. In some embodiments, data nodes 110B, 110C, 110D are organized into racks where all nodes are connected via a network switch. The network speed between nodes within a rack can be faster than the network speed between nodes in different racks. The Hadoop cluster may take that fact into account when assigning tasks. A DataNode service 170 runs on each data node for responding read and write requests for blocks. The

DataNode services 170 also respond to requests from the metadata node for creating, deleting and replicating blocks.

[0044] In some embodiments, the data nodes 110B, 110C, 110D send periodic heartbeat messages including block reports to the metadata node 110A. The metadata node 110A validates its block mapping and other file system metadata using the periodic block reports. [0045] When a client 1 30A or 1 30B seeks to write a file to the Hadoop cluster 1 00, the client sends a file creation request to the metadata node 1 1 0A. The metadata node 1 1 0A responds to the client with the identity of one or more assigned data node(s) and the destination locations of the blocks of the files. The client sends data blocks of the fi le to the assigned data node(s); and the cluster can repl icate the data blocks on one or more data nodes. Once all blocks are sent, the metadata node records the file creation i n its metadata, including the Fsl mage file.

[0046] A Hadoop cluster works as a parallel data processing engine based on a framework called MapReduce. The Hadoop cluster includes a JobTracker to implement the MapReduce fu nction . As il lustrated in FIG . 2, a JobTracker can be i mplemented as a dedicated server (JobTracker node 21 0A) within a Hadoop cluster 200. The Hadoop cluster 200 includes two racks 242 and 244 interconnected via a switch 226. Rack 242 includes JobTracker node 21 0A, metadata node 21 0B, data nodes 21 0C-21 0D and a switch 222 to interconnect the nodes within rack 242. Rack 244 includes data nodes 21 0E-21 0H and a switch 224 to interconnect the nodes within rack 242. In some other embodiments, a JobTracker can be im plemented as a service that shares the same metadata node with a NameNode service. The metadata node 21 0B (also referred to as name node) runs the NameNode service to track where across the cluster the data is kept. The JobTracker node 21 0A, which is dedicated to control MapReduce jobs, receives a request from a client 230 to launch a MapReduce job. Once the MapReduce job (also referred to as a MapReduce appl ication or a

MapReduce task) is subm itted to the JobTracker 21 OA, JobTracker 21 OA identifies input and output files and/or directories in the H DFS for the job . The input file for the MapReduce task can incl ude a plurality of in put file blocks that contain in put data for the MapReduce task. The JobTracker 21 0A uses knowledge of the i nput file blocks (including physical quantity of blocks and where the blocks are located) to decide how many subordinate tasks wil l be created . The MapReduce application is copied to every processi ng node where the input file blocks are present. For each assigned processing node , JobTracker 21 OA creates at least one subordinate task. On each assigned processing node, a TaskTracker service mon itors the status of the su bordinate task on that node and reports the status and intermediate output back to the JobTracker. The Hadoop cluster 200 distributes the subordinate tasks based on the knowledge of the fi le blocks. Therefore, instead of moving storage to the processing locations, a Hadoop cluster moves the processi ng tasks to the storage locations.

[0047] Although the nodes 21 0A-21 0H are il lustrated as single units in F IG . 2, each node can have a distributed architecture. For example, a node can be designed as a combination of multiple computers, which may be physically separate from each other and which may communicate with each other over a physical interconnect. Such an architecture allows convenient scal ing , such as by deploying computers capable of

communicating with each other through the interconnect.

[0048] I n one embodiment, a distributed query engi ne is built on top of a Hadoop cluster running Hadoop Distributed File System (H DFS) and H ive data warehouse. FIG . 3 illustrates the relationships among the distri buted query engine 300, the Hive data warehouse, the H DFS and the storage cluster. The distributed query engine 300 is bui lt on top of the H ive data warehouse and the H DFS, which in turn rely on the storage cluster to operate. A Hadoop cluster 31 0 including a plu rality of nodes 31 2 that are responsible for storing the massive amount of data. The Hadoop cl uster 31 0 further includes a metadata node 31 4. A Hadoop Distributed File System (H DFS) 320 is running on the Hadoop cluster 31 0 to distribute and manage the data among the nodes 31 2. Data Node services 322 are running on the nodes 31 2 to manage the local data store in the nodes 31 2. Data and files are split into blocks that are stored in the nodes 31 2 of the Hadoop cluster 31 0. A NameNode service 324 is running on the metadata node 31 4 to provide metadata services within the Hadoop cluster, including maintaining the file system namespace and controlling access by external clients.

[0049] A H ive data warehouse system 330 is built on top of the Hadoop cluster 31 0 and the H DFS 320. The H ive data warehouse system 330 works as a database interface. The H ive data warehouse system 330 supports data queries expressed in an SQL-like declarative language, H iveQL. Instead of relying on the map-reduce scheme of the Hadoop cluster to process the database queries for the Hive system , the

discontinued query engine includes a plurality of worker nodes 342 to process the database queries in a parallel fashion . The distributed query engine 300 further includes a gateway 348. I n one embodiment, the worker nodes 342 are im plemented as services running on the nodes 312 of the Hadoop cluster 31 0. In another em bodiment, the worker nodes 342 are implemented as dedicated servers interconnected with the nodes 31 2 of the Hadoop cluster 31 0.

[0050] I n some embodiments, the worker nodes 342 are responsible to compi le the partial tasks expressed in HiveQL into instructions that the DataNode services 322 of the H DFS 320 can execute.

[0051 ] The distributed query engi ne receives a query task from a client and splits the query task into multiple partial tasks. F IG . 4 i llustrates a high level block diagram of an example distributed query engine 400.

Operators of the distributed query engine 400 can provide query task via input interface 486 of a cl ient 480. I n one embodiment, the input interface 486 includes both command line interface 482 and graph ic interface 484. Using the command line interface 482, an operator can provide the query task as a program directly expressed i n a database query language, such as SQL or H iveQL. By using the graphic interface 484, an operator can provide the query task by working with the graphic interface elements of 484. I n one embodiment, the graphic interface 484 is implemented as an input web page . An operator can provide the query task by interacting with elements on the input web page, selecting options, and entering input data. The graphic interface 484 can translates the operator's selections and in puts into a correspondi ng program expressed in a database query language . The input interface 486 transmits the program , received from either the command li ne interface 482 or graphic i nterface 484, to a gateway 41 0 of the distributed query engine 400 as the query task. [0052] The gateway 41 0 receives the query task from the client 480 and parses the query task. The gateway 41 0 sends an inqui ry to the H ive Metastore 440 based on the query task. The Hive Metastore 440 returns table metadata and H DFS file identifications to the gateway 41 0 for the data upon which the query task needs to run . Then the gateway 41 0 retrieves the locations of the corresponding H DFS blocks from the H DFS Name Node 460 based on the H DFS file identifications.

[0053] I n one embodiment, the gateway 41 0 divides the query task into a plurality of partial queries based on the corresponding H DFS blocks. The gateway 41 0 assigns each individual partial query to be executed on one H DFS block within the correspondi ng H DFS blocks. I n other embodiments, the gateway 41 0 can divide the query task i nto partial queries in other ways, as may be contemplated by a person of ordinary skill in the art.

[0054] The gateway 41 0 sends each of the partial queries to a worker 41 2 for local processing. I n one embodiment, workers 41 2 are overlaid on the Hadoop cluster that stores the H ive table data. Each work 41 2 runs as a service on a Hadoop cluster node 432. The partial queries are

generated such that each worker 41 2 is responsible for a partial query to process data store on the node 432 which that particular worker 41 2 runs. The worker 41 2 directly contacts a DataNode service 422 which runs on the same cluster node 432 as the worker 41 2. The worker 41 2 is able to ach ieve low-latency data reads for the partial query by requesting data within a single cluster node 432.

[0055] The cluster node 432 can use remote procedure cal l (RPC) frameworks to facilitate the implementation of the services. For example, in one embodiment, the cluster node 432 use an RPC framework such as

Apache Th rift framework to define and create the worker service 41 2 as a highly scalable and high performance server service.

[0056] I n one embodiment, the worker nodes 342 are implemented as services running on the nodes 31 2 of the Hadoop cluster 31 0. I n another embodiment, the worker nodes 342 are implemented as dedicated servers interconnected with the nodes 31 2 of the Hadoop cluster 31 0. [0057] The workers 41 2 periodically retu rn status updates (referred to as "heartbeats") to the gateway 41 0 indicating the process of the partial query processing. I n one embodiment, if there is an assigned worker which stops returni ng heartbeats or shows no progress, the gateway 41 0 determ ines that the worker has failed and reschedules the partial query to another worker. Each worker 41 2 scans data related to the respective partial query that stored on one or more cluster nodes 432, and generates an

intermediate result for the partial query. I n one embodiment, the worker 41 2 processes the partial query entirely in the memory of the cluster node on which the worker 41 2 runs. The worker 41 2 stores the intermediate result in its memory. At the end of the processing of the partial query, the worker 41 2 sends the intermediate result to the gateway 41 0. I n some embodiments, the worker 412 sends the i ntermediate result via RPC cal ls such as Apache Thrift calls.

[0058] The gateway 41 0 receives all intermed iate results from the workers 41 2 and combines the intermediate results into a query result as an answer for the query task. Then the gateway 41 0 returns the query result to the client 480. I n one embodiment, the cl ient 480 optional ly displays the query result on a display component.

[0059] A MetaCache 41 4 runs on the gateway 41 0 to cache both the H ive table-level and the H DFS file-level metadata to decrease query latency. I n some embod iments, the MetaCache 41 4 can be implemented as a standalone server interconnected with the gateway 41 0. The

MetaCache 41 4 can keep the cached data from previous queries. For example, when an operator is interactively analyzi ng on data of a H ive table, the operator ru ns mu ltiple consecutive queries on the same Hive table. By preserving the cached data from previous queries, the

MetaCache 41 4 can reuse the cached metadata instead of fetching the metadata repetitively from the Hive Metastore 440 and the H DFS

NameNode 460.

[0060] The MetaCache 41 4 has high cache hit rates because data in a typical Hive table is written once and read many times without further changes. I n one embodiment, the MetaCache 41 4 can retrieve a real-time feed of audit logs of the H ive system to invalidate entries in the cached data for partitions in the H ive system that may have been changed by H ive queries or other operations. I n another embodiment, the MetaCache 414 automatically pu rges entries in the cached data that have not been queried for a predeterm ined time period (e.g . one hour). I n doing so, the

MetaCache 41 4 prevents arbitrary growth in memory usage and mi nim izes cache errors.

[0061 ] The workload of each cluster node 432 on wh ich a worker 41 2 runs can be different. Cluster nodes 432 and worker services 41 2 can also fail for various reasons. While the gateway 41 0 can receive intermediate results from a majority of workers 41 2 in a reasonable time period , there are workers 41 2 that fail to deliver the intermediate result due to the node or service failures or delays. These workers either report rates of progress below a predetermined percentage after a predetermined time period since the partial queries distribution , or si mply fail to respond the progress to the gateway 41 0. These workers are identified as straggling workers. Once the gateway 41 0 identifies the straggling workers, the gateway 41 0 sends messages to the straggling workers to cancel assignment of the partial queries. For each unfinished partial query for wh ich straggling workers failed to deliver the intermediate results, the gateway 41 0 further divides the partial query i nto a plurality of subordinate partial queries, and assigns the subordinate partial queries to some of the workers 41 2. I n one embodiment, the gateway 41 0 determines the assignment of the

subord inate partial queries based on the current workload of the workers 41 2. I n other embodiments, the gateway 41 0 can determine the

assignment in other ways, as may be contemplated by a person of ordinary ski ll in the art. Th is add itional parallelized process speeds up the retries of the unfinished partial queries and thus decreases the query latency caused by straggling workers.

[0062] FIG . 5 illustrates a sample process for identifying straggl ing worker nodes and further d ividing the partial queries. At step 502, a gateway of the distributed query engine receives a database query from a client device, for a database containing data stored in a distributed storage cluster which has a plurality of data nodes. At step 504, the gateway divides the database query into a plurality of partial queries. Then at step 506, the gateway sends each of the partial queries to a respective worker node of a plurality of worker nodes. Each worker node can be a service running on a data node of the distributed storage cluster.

[0063] At step 508, the gateway server identifies a straggling worker node. The gateway further divides a partial query that is assigned to the straggling worker node into a plurality of subordi nate partial queries, and assigns the plurality of subordi nate partial queries to some of the plurality of worker nodes. I n one embodiment, the gateway identifies a straggling worker node by monitoring heartbeat messages that the worker nodes periodically sends. The straggling worker node is identified when

heartbeat messages from the straggling worker node are not received for a predetermined time period . I n another embodi ment, the straggling worker node is identified when a heartbeat message from the straggling worker node is received; where the heartbeat message includes a number representing the status of a partial query processing by the straggling worker node that is below a threshold value.

[0064] At step 51 0, the gateway retrieves a plurality of intermediate results for the partial queries from the worker nodes. Each intermediate result is processed by a respective worker node of the worker nodes by scanning related data stored in a data node on wh ich the perspective worker node runs. At step 51 2, the gateway generates a query result based on the plurality of intermed iate resu lts.

[0065] I n one embodiment, the distributed query engine can be

compatible with H ive system's data formats and metadata , and can support a subset or an entire set of HiveQL language. And H iveQL is a declarative language similar to SQL. H iveQL does not need to strictly follow the standard of SQL, and offers extensions not originally specified in SQL. For example, the distri buted query engi ne can support filters, aggregates, top-k, percentiles, sub-queries in FROM clauses, U N ION ALL and user-defined functions. [0066] I n one embodiment, the distributed query engine support

TABLESAM PLE clause can be used to explicitly lim it the amount of input data that is scanned . I n another embodiment, the distributed query engine support W ITH clause allows operators to write more readable queries by declaring variables for complex expressions that are used multiple times in the same query. The W ITH clause can also provide a way for the operator to specify hints to the optim izer such that it eval uates common

sub-expressions only once during run time.

[0067] I n some embodiments, the distributed query engine can supply an approximate query resu lt before the entire query processing is finished . The distributed query engine can automatically supply the approximate query result after a predetermined time period since the initial query in put, or when the processing of the query satisfies a predetermined condition (e.g . a certain number of workers fail). The distributed query engine can also supply the approximate query result i n response to an operator instruction . For example, an operator waiti ng for the query result can input Ctrl-C to instruct the d istributed query engine to stop the query processing. U pon receiving the instruction , the distributed query engi ne stops the query processing and returns an approximate query result. In one embodiment, the distributed query engine further supplies a percentage indicator to indicate the percentage of data scanned for the approximate query result. I n one em bodiment, the distributed query engine returns the approximate query resu lt and continues the query processing for an exact query result (i .e. with a percentage indicator of 1 00%).

[0068] Particularly for queries to explore data instead of write or change data, approximate query results can be sufficient for the operator's analysis. Run time errors l ike node fai lures, corrupted input data, or even a user kill ing their queries can be treated as situations where all input data was not scanned . When there is a failure, the distributed query engine can instantly retu rn an approximate query result based on the processi ng of the partial queries so far, instead of simply returni ng an error message. In one embodiment, the distributed query engine returns the approximate query result along with a percentage ind icator. [0069] I n another embodiment, an operator can also specify in his query task that an exact query result is required . I n these cases, a run ti me error can be returned if the query processing fails.

[0070] The distributed query engine uses one-pass algorithms for aggregations and stores all intermediate results in memory. The sizes of the intermediate results and the final query result can be relatively small . The distributed query engine's ability of returning an approximate query result can further reduces the size. For example, if the query is to sort data records by a specified column (e.g . ORDER BY clause), the

distributed query engine can generate approximate answer by letting workers only scan a portion of the related data records. Simi larly, the distributed query engine can also generate approximate answers for types of queries such as counting distinct entries and calculating percentile.

[0071 ] For example, in one em bod iment, an operator of the distributed query engine can i nput a query task instructi ng a count of the number of distinct values of a country colum n from a specific set of data records. The distributed query engine receives the query, divides the query i nto partial queries, and assigns workers to process these partial queries. After 20 seconds from the time the distributed query engine starts the task, the operator terminates the task by pressing CTRL-C on a keyboard of a client device. Upon receiving the termi nation instruction , the d istributed query engine immediately instructs the assigned workers to return approximate intermediate results and in turn returns an approximate result by combining these approxi mate intermed iate resu lts. The approximate result is returned to a client device of the operator. The client device can further receive information about the total processing ti me , num ber of workers used , rows of the data records that have been scanned , amount of data scanned , amount of data to be scanned for exact result, percentage of data scanned, and/or number of fail ures.

[0072] I n another embod iment, the distributed query engine can automatically return an approximate query result after the number of failures exceeding a predetermined threshold value. I n yet another embodiment, the operator specifies in the query task that an exact result is requi red . The distributed query engine will keep processing u ntil all related data records are scanned by assigned workers. An exact query result can be returned to a user device of the operator. The client device can further receive information about the total processing time, number of workers used, rows of the data records that have been scanned , amount of data scanned, percentage of data scanned (i .e. 1 00%), and/or number of failures.

[0073] FIG . 6 ill ustrates a sample process for approximate processing of a database query. At step 602, a gateway of the distributed query engine receives a database query from a client device, for a database containi ng data stored i n a distributed storage cluster which has a plurality of data nodes. At step 604, the gateway divides the database query i nto a plurality of partial queries. Then at step 606, the gateway sends each of the partial queries to a respective worker node of a plurality of worker nodes. Each worker node can be a service running on a data node of the distributed storage cluster. At step 608, after send ing the partial queries, the gateway can instruct the worker nodes to immediately retu rn

approximate query resu lts. I n one embodiment, immediately returning approximate query results means returning the results within a very short time period such as one second . The instruction can be triggered by various events. For example, the gateway may receive an instruction from the client device to terminate the processing of the database query; or the gateway may decide to return an approximate query result automatically when an exact query resu lt is not available after a predeterm ined time period . Therefore, the approximate result can be requested manually by the client device, or automatically triggered by the distributed query engine without user intervention .

[0074] At step 61 0, after the instruction for approximate intermediate results, the gateway retrieves a plurality of approximate intermediate results for the partial queries from the worker nodes. Each approximate intermediate result is processed by a respective worker node of the worker nodes by scanning a portion of related data stored in a data node on which the perspective worker node runs. Upon receivi ng the approxi mate intermediate results, at step 61 2, the gateway generates an approximate query result based on a plu rality of approximate intermediate results.

Then at step 61 4, the gateway of the distributed query engine returns the approximate query result. I n one embodiment, the approximate query result is returned along with a percentage indicator to the client device. The percentage indicator suggests the percentage of related data stored in the data nodes that has been scanned for the query result.

[0075] I n addition to the above illustrated advantages, the techniques presented herein present additional advantages as discussed below.

[0076] The distributed query engine significantly reduces the latencies of queries on data stored in data storage clusters such as a Hadoop cluster, as compared to the query latencies by using a Hive system alone.

Operators or users of the distributed query engine can conduct ad hoc queries with smal l waiting time periods. The distributed query engine can be utilized in different situations. For example, without the distributed query engine, the operators or analysts need to explicitly load data from H ive into a database such as MySQL or Oracle and then fetch data from the database to drive web-based data analysis reports. Using the distributed query engine, the operators can fetch data directly from the H ive system to generate the web-based data analysis report.

[0077] I n one embodiment, when an operator is authori ng a query task usi ng, for example, the graph ic interface 484 shown in FIG . 4, the distributed query engine can fetch data samples to show the operator a preview of the data on the graphic interface 484. I n another em bod iment, the distributed query engine can index and pin popular data sets in memory to further improve query latencies.

[0078] The techniques presented herein provide a low latency

distributed query engine that can be built on top of a data storage cluster. The distributed query engine can be compatible with existing data and metadata of a H ive system . The distributed query engine can be used to drive data analysis reports without the need for pipelines that load data into other databases (e.g . MySQL or Oracle) for ad-hoc analysis. [0079] F IG . 7 is a high-level block diagram showi ng an example of the architectu re of a com puter node, which may represent any cluster node described herein. The node 700 includes one or more processors 71 0 and memory 720 coupled to an interconnect 730. The interconnect 730 shown in Figure 7 is an abstraction that represents any one or more separate physical buses, point to point connections, or both connected by

appropriate bridges, adapters, or controllers. The interconnect 730, therefore, may i nclude, for example, a system bus, a Peripheral

Component Interconnect (PCI ) bus or PC I-Express bus, a HyperTransport or industry standard architecture (ISA) bus, a small computer system interface (SCS I ) bus, a universal serial bus (USB), I IC (I2C) bus, or an I nstitute of Electrical and Electronics Engi neers (I EE E) standard 1 394 bus, also called "Firewire".

[0080] The processor(s) 71 0 is/are the central processing unit (CPU ) of the storage controller 700 and , thus, control the overal l operation of the node 700. I n certain embodiments, the processor(s) 71 0 accom plish this by executing software or firmware stored in memory 720. The processor(s) 71 0 may be, or may include, one or more programmable general-purpose or special-purpose microprocessors, digital signal processors (DSPs), programmable controllers, application specific integrated circuits (AS ICs), programmable logic devices (PLDs), trusted platform modules (TPMs), or the like, or a combination of such devices.

[0081 ] The memory 720 is or includes the main memory of the node 700. The memory 720 represents any form of random access memory (RAM), read-only memory (ROM), flash memory, or the like, or a combination of such devices. In use, the memory 720 may contain a code 770 containing instructions according to the techniques disclosed herein .

[0082] Also connected to the processor(s) 71 0 through the interconnect 730 are a network adapter 740 and a storage adapter 750. The network adapter 740 provides the node 700 with the ability to communicate with remote devices, over a network and may be, for example, an Ethernet adapter or Fibre Channel adapter. The network adapter 740 may also provide the node 700 with the abi lity to communicate with other nodes within the cluster. In some embodiments, a node may use more than one network adapter to deal with the communications within and outside of the cluster separately. The storage adapter 750 allows the node 700 to access a persistent storage, and may be, for example, a Fibre Channel adapter or SCS I adapter.

[0083] The code 770 stored in memory 720 may be im plemented as software and/or firmware to program the processor(s) 71 0 to carry out actions described above . I n certain embodiments, such software or firmware may be initial ly provided to the node 700 by downloading it from a remote system th rough the node 700 (e.g . , via network adapter 740).

[0084] The techniques introduced herein can be implemented by, for example, programmable circuitry (e.g . , one or more microprocessors) programmed with software and/or firmware, or entirely i n special-purpose hardwired circuitry, or in a combination of such forms. Special-pu rpose hardwired ci rcuitry may be in the form of, for example , one or more application-specific integrated circuits (AS ICs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), etc.

[0085] Software or firmware for use in i mplementing the techniques introduced here may be stored on a machi ne-readable storage medium and may be executed by one or more general-purpose or special-purpose programmable microprocessors. A "machine-readable storage medium", as the term is used herein , includes any mechanism that can store information in a form accessi ble by a machine (a machine may be, for example, a computer, network device, cellular phone, personal digital assistant (PDA), man ufacturing tool , any device with one or more

processors, etc. ). For example, a machine-accessible storage medium includes recordable/non-recordable media (e.g . , read-only memory (ROM ); random access memory (RAM); magnetic disk storage media ; optical storage media; flash memory devices; etc. ), etc.

[0086] The term "logic", as used herein , can incl ude, for example, programmable ci rcuitry program med with specific software and/or firmware, special-purpose hardwi red circu itry, or a combination thereof. [0087] I n addition to the above mentioned examples, various other modifications and alterations of the invention may be made without departing from the invention. Accordingly, the above disclosure is not to be considered as lim iting and the appended clai ms are to be interpreted as encompassing the true spirit and the entire scope of the invention .

Claims

1 . A system comprising :

a gateway server configured to generate a plu rality of partial queries from a database query for a database containi ng data stored in a distributed storage cluster that has a plurality of data nodes, and to construct a query result based on a plurality of intermediate results; and

a plurality of worker nodes, wherein each worker node of the plurality of worker nodes is configured to process a respective partial query of the plurality of partial queries by scanning data related to the respective partial query and stored on at least one data node of the distributed storage cluster, and wherein each worker node of the plurality of worker nodes is further configured to generate an intermediate result of the pl urality of intermediate resu lts that is stored in a memory of that worker node .

2. The system of claim 1 , wherein each worker node of the plurality of worker nodes is further configured to process the respective partial query of the plurality of partial queries by scanning a portion of the data related to the respective partial query that is stored on the at least one data node of the distributed storage cl uster and to generate an approximate intermediate result that is stored in the memory of that worker node;

wherein preferably the gateway server is fu rther configured to construct an approximate query result based on at least one approximate intermediate result.

3. The system of claim 1 or 2, wherein the gateway server is further configu red to construct an approximate query result based on a portion of the plu rality of intermediate results.

4. The system of any of claims 1 to 3, wherein the gateway server is further configured to identify a straggling worker node, further divide a partial query that is assigned to the straggling worker node into a plural ity of subordinate partial queries, and assign the plurality of subordinate partial queries to some of the plurality of worker nodes, wherei n the straggling worker node is a worker node that either fails to report a rate of progress to the gateway server or reports the rate of progress below a predetermined value after a predetermined time period to the gateway server.

5. The system of any of claims 1 to 4 , wherein each worker node of the plural ity of the worker nodes is a service running a respective data node with in the distributed storage cluster.

6. The system of any of claims 1 to 5 , further comprising:

a metadata cache configured to cache table level metadata of the database and file level metadata of the d istributed storage cluster; wherei n preferably the metadata cache is configured to retain cached metadata from a previous database query for the database query.

7. The system of any of claims 1 to 6 , wherein each worker node of the plurality of the worker nodes periodical ly sends heartbeat messages to the gateway server to report status of a partial query processing by that worker node.

8. The system of any of claims 1 to 7, wherein the gateway server is further configured to receive an instruction from a client device to return an approximate query result or term inate a processing of the database query.

9. The system of any of claims 1 to 8, wherein the gateway server is further configured to instruct the worker nodes to immediately return approximate intermediate results, and to return an approximate query result based on the approxi mate intermediate results to a client device.

1 0. The system of any of claims 1 to 9, wherein the database query

incl udes a request for an approxi mate query result.

1 1 . The system of any of claims 1 to 1 0, wherein the query result is

accompanied by an indication of a portion of related data stored in the data nodes that has been scanned for the query result.

1 2. The system of any of claims 1 to 1 1 , wherein the database is a Hive data warehouse system and the distributed storage cluster is a Hadoop cluster.

1 3. A method wherein a system according to any of the claims 1 to 1 2 is used .

1 4. A method comprising:

dividing the database query into a plurality of partial queries;

sending each of the partial queries to a respective worker node of a plurality of worker nodes, wherein each worker node is a service running on a data node of the distributed storage cluster;

retrieving a plurality of intermediate results for the partial queries from the worker nodes, wherein each intermediate result is processed by a respective worker node of the worker nodes by scanning related data stored in a data node on which the perspective worker node runs; and

generating a query result based on the plurality of intermediate results.

15. The method of claim 14, further comprising:

returning the query result along with a portion indicator to the client device, wherein the portion indicator indicates the portion of related data stored in the data nodes that has been scanned for the query result.

1 6. The method of claim 1 4, further comprising :

instructing the worker nodes to immediately return approximate query results;

and wherein the step of retrieving comprises:

retrievi ng a plurality of approximate intermediate results for the partial queries from the worker nodes, wherein each approximate intermediate result is processed by a respective worker node of the worker nodes by scanni ng a portion of related data stored in a data node on which the perspective worker node runs.

1 7. The method of claim 1 4, further comprising:

for each partial query, retrieving metadata regarding which data node stores data related to the partial query;

and the step of sending comprises:

1 8. A method comprising :

receiving a database query from a cl ient device, for a database containing data stored in a distributed storage cl uster having a plural ity of data nodes;

dividing the database query into a plurality of partial queries;

sending each of the partial queries to a respective worker node of a plural ity of worker nodes, wherein each worker node is a service running on a data node of the distributed storage cluster; identifying a straggling worker node, dividing a partial query that is assigned to the straggling worker node into a plu rality of subordinate partial queries, and assigning the plurality of subordinate partial queries to some of the plural ity of worker nodes;

retrievi ng a plurality of intermediate results for the partial queries from the worker nodes, wherein each intermediate result is processed by a respective worker node of the worker nodes by scanni ng related data stored in a data node on which the perspective worker node runs; and

generating a query result based on the plurality of intermediate results.

The method of claim 1 8, wherein the step of identifying comprises: identifying a straggling worker node by mon itoring heartbeat messages that the worker nodes periodically send , wherein the straggling worker node is identified when heartbeat messages from the straggli ng worker node are not received for a predetermined time period , or when a heartbeat message from the straggling worker node is received and the heartbeat message including a number representing a status of a partial query processing by the straggling worker node that is below a th reshold value.