Building Highly Available Database Applications with Geronimo and Derby

Transcription

1 Building Highly Available Applications with Geronimo and Emmanuel Cecchet Principal architect - Continuent Sequoia is an open source software (Apache v2 License) hosted by Continuent.org. Sequoia is the continuation of the C-JDBC project hosted by the ObjectWeb consortium ( 1

2 Motivations tier should be scalable highly available without modifying the client application database vendor independent on commodity hardware Internet JDBC JDBC JDBC 2

3 Single database issue Clients connect to the application server Application server builds web pages with data coming from the database Application server clustering solves application server failure outage causes overall system outage Application servers Internet Disk 3

4 Disk replication/clustering Eliminates the single point of failure (SPOF) on the disk Disk failure does not cause database outage outage problem still not solved Application servers Internet Disks 4

5 Master/slave replication Lazy replication at the disk or database level No scalability Data lost at failure time System outage during failover to slave Failover can take several hours Application servers Master Disks log shipping hot standby Internet Slave 5

6 Scaling the database tier Master-slave replication Cons failover time/data loss on master failure read inconsistencies scalability Web frontend App. server Master Internet MySQL replication uses a master/slave replication scheme. Writes must be directed to the master node and updates are lazily replicated to slave nodes. 6

7 Scaling the database tier Atomic broadcast Cons atomic broadcast scalability no client side load balancing heavy modifications of the database engine Internet Atomic broadcast Queries are sent atomically in the same total order to all replicas. Optimizations may include executing the write query at one node and propagate only the modified rows to the other nodes. This requires heavy modifications of the database engine and makes locking management very tedious. 7

8 Scaling the database tier SMP Cons Cost Scalability limit Web frontend App. server Internet Well-known database vendor here Well-known hardware + database vendors here This is the most common solution in commercial setups: large SMP machine and prohibitive DB license costs 8

9 clustering with shared disk Multiple database instances share the same disk Disk can be replicated to prevent SPOF on disk No dynamic load balancing failure not transparent to users (partial outage) Failed instance leaves incorrect data in the disks Manual failover + manual cleanup needed Application servers s Disks Internet 9

10 Sequoia competitive advantages High availability with fully transparent failover Scalability with dynamic load balancing Commodity hardware No modification to existing client application vendor independent Heterogeneity support Platform independent 10

11 Outline RAIDb Sequoia Geronimo/Sequoia/ Scalability High availability 11

12 RAIDb concept Redundant Array of Inexpensive s RAIDb controller gives the view of a single database to the client balance the load on the database backends RAIDb levels offers various tradeoff of performance and fault tolerance RAIDb is to databases what RAID is to disks. 12

13 RAIDb levels RAIDb-0 partitioning no duplication and no fault tolerance at least 2 nodes SQL requests RAIDb controller table 1 table 2 & 3 table... table n-1 table n 13

14 RAIDb levels RAIDb-1 mirroring performance bounded by write broadcast at least 2 nodes SQL requests RAIDb controller Full DB Full DB Full DB Full DB Full DB 14

15 RAIDb levels RAIDb-2 partial replication at least 2 copies of each table for fault tolerance at least 3 nodes SQL requests RAIDb controller Full DB table x table y table x & y table z 15

16 RAIDb levels composition RAIDb-1-0 no limit to the composition deepness RAIDb-0 controller SQL requests RAIDb-1 controller RAIDb-0 controller table w table x & y table z table w table y table x & z RAIDb-0 controller table w table x table y table z 16

17 Outline RAIDb Sequoia Geronimo/Sequoia/ Scalability High availability 17

18 Sequoia overview Middleware implementing RAIDb 100% Java implementation open source (Apache v2 License) Two components generic JDBC (Sequoia ) Sequoia Controller Read-one, Write all approach provides eager (strong) consistency Supports heterogeneous databases 18

19 Sequoia architectural overview Sequoia JDBC Driver JVM Sequoia controller JVM Sequoia can be seen as a JDBC proxy between the application and the database native. 19

20 Sequoia vs C-JDBC License Bug tracking Hosting Write scalability C-JDBC LGPL GForge ObjectWeb Single threaded Sequoia APL v2 JIRA Continuent.org Multi threaded Group communication Tribe/JGroups Hedera/JGroups C/C++ support Mgt console - Text Carob Text + Oak Eclipse plugin Sequoia is the continuation of the C-JDBC project ( This slide shows the main differences between Sequoia and C-JDBC. 20

21 Inside the Sequoia Controller Client application (Servlet, EJB,...) Sequoia Client application (Servlet, EJB,...) Sequoia Client application (Servlet, EJB,...) Sequoia Sequoia Controller Virtual database 1 Virtual database 2 HTTP RMI... JMX administration console Configuration Monitoring JMX Server Authentication Recovery Log Request Scheduler Query result cache Load balancer Authentication Recovery Log Request Scheduler Query result cache Load balancer dumps management Checkpointing service DB2 JDBC DB2 JDBC DB2 JDBC MS SQL JDBC Oracle JDBC DB2 DB2 DB2 MS SQL Oracle The Sequoia controller is a Java application. It can be co-located with the or the database or run on a separate machine. 21

22 Virtual Java client program (Servlet, EJB,...) Sequoia Virtual database Authentication Recovery Log Request Scheduler Request Cache Load balancer Gives the view of a single database Establishes the mapping between the database name used by the application and the backend specific settings s can be added and removed dynamically s can be transferred between controllers configured using an XML configuration file DB2 JDBC DB2 JDBC DB2 JDBC DB2 DB2 DB2 22

23 Authentication Java client program (Servlet, EJB,...) Sequoia Virtual database Authentication Recovery Log Request Scheduler Request Cache Load balancer Matches real login/password used by the application with backend specific login/ password Administrator login to manage the virtual database DB2 JDBC DB2 JDBC DB2 JDBC DB2 DB2 DB2 23

24 Scheduler Java client program (Servlet, EJB,...) Sequoia Virtual database Authentication Recovery Log Request Scheduler Request Cache Load balancer Manages concurrency control Assigns unique id to requests and transactions Ensure serializable transactional order Relies on database (row-level) locking DB2 JDBC DB2 JDBC DB2 JDBC DB2 DB2 DB2 24

25 Request cache DB2 JDBC Java client program (Servlet, EJB,...) Sequoia Virtual database Authentication Recovery Log Request Scheduler Request Cache Load balancer DB2 JDBC DB2 JDBC 3 optional caches tunable sizes parsing cache parse request skeleton only once INSERT INTO t VALUES (?,?,?) metadata cache column metadata fields of a request result cache caches results from SQL requests tunable consistency fine grain invalidations optimizations for findbypk requests DB2 DB2 DB2 25

26 Result cache Cache contains a list of SQL ResultSet Policy defined by querypattern Policy 3 policies EagerCaching: variable granularities for invalidations RelaxedCaching: invalidations based on timeout NoCaching: never cached RUBiS bidding mix No cache Coherent Relaxed with 450 clients cache cache Throughput (rq/min) Avg response time 801 ms 284 ms 134 ms CPU load 100% 85% 20% C-JDBC CPU load - 15% 7% Eager caching accepts database, table, column and column unique granularities. Basically, every request is parsed and C-JDBC checks which tables or rows are needed to compute the result. When a write queries is executed, it invalidates all results that required the data that has been invalidated. The finer grain invalidation requires more extensive parsing (more cpu) but have better hit ratio. Relaxed caching can further enhance the cache performance by relaxing the consistency of queries that can accept a given staleness (for example a page whose content is only refreshed every minute). 26

27 Load balancer 1/2 DB2 JDBC Java client program (Servlet, EJB,...) Sequoia Virtual database Authentication Recovery Log Request Scheduler Request Cache Load balancer DB2 JDBC DB2 JDBC RAIDb-0 query directed to the backend having the needed tables RAIDb-1 read executed by current thread write executed in parallel by a dedicated thread per backend result returned if one, majority or all commit if one node fails but others succeed, failing node is disabled RAIDb-2 same as RAIDb-1 except that writes are sent only to nodes owning the updated table DB2 DB2 DB2 RAIDb-0 is for partitioning (tables distributed on the different backends). RAIDb-1 is for full mirroring. RAIDb-2 is for partial replication. 27

28 Load balancer 2/2 Java client program (Servlet, EJB,...) Sequoia Virtual database Authentication Recovery Log Request Scheduler Request Cache Load balancer Static load balancing policies Round-Robin (RR) Weighted Round-Robin (WRR) Least Pending Requests First (LPRF) request sent to the node that has the shortest pending request queue efficient even if backends do not have homogeneous performance DB2 JDBC DB2 JDBC DB2 JDBC DB2 DB2 DB2 28

29 DB2 JDBC Java client program (Servlet, EJB,...) Sequoia Virtual database Authentication Recovery Log Request Scheduler Request Cache Load balancer DB2 JDBC DB2 JDBC Sequoia JDBC provides transparent connection pooling pooling for a backend no pooling blocking pool non-blocking pool dynamic pool pools defined on a per login basis resource management per login dedicated connections for admin DB2 DB2 DB2 29

30 Recovery Log Java client program (Servlet, EJB,...) Sequoia Virtual database Authentication Recovery Log Request Scheduler Request Cache Load balancer Checkpoints are associated with database dumps Record all updates and transaction markers since a checkpoint Used to resynchronize a database from a checkpoint Store log information in a database Can be re-injected in a Sequoia cluster for fault tolerance DB2 JDBC DB2 JDBC DB2 JDBC DB2 DB2 DB2 30

31 Functional overview establishment Read request Write request Failure handling Recovery Summary 31

32 establishment Sequoia JDBC URL jdbc:sequoia://controller1[:port1],controller2[:port2]/vdbname Driver connection to a controller according to a policy defined in the URL (random, round-robin, ordered, ) Controller connections to a physical database use a connection pool defined per backend All queries on a connection go to the same controller but controller load balance queries on the underlying backends 32

33 Sequoia read request connect execute SELECT login, connect password * mydb FROM t Client application Client application ordering (Servlet, EJB,...) (Servlet, EJB,...) Sequoia Sequoia Sequoia Controller update Virtual database 1 (if available) cache RR, WRR, Authentication exec get from connection LPRF, pool Request Scheduler Recovery Query result cache Log Load balancer JDBC JDBC JDBC The connection is first authentified by the authentication manager. Then SELECT requests are scheduled, checked again the cache (a cache hit immediately returns) and then given to the load balancer that chooses a backend to execute the query. Upon completion, the cache is updated and the result is returned to the client. 33

34 Write request Query broadcast using total order between controllers s execute query in parallel (possibly asynchronously) Automatically disable backends that fail if others succeed Update recovery log (transactional log) asynchronously Non-conflicting transactions execute in parallel table-based locking used as a hint for detecting conflicts conflicting queries execute in a single-threaded queue stored procedures execute one at a time 34

35 Controller replication Client application (Servlet, EJB,...) p/cluster Client application (Servlet, EJB,...) p/cluster Client application (Servlet, EJB,...) p/cluster jdbc:sequoia://node1,node2/mydb Total order reliable multicast p/cluster Controller Distributed Request Virtual database 1 Authentication p/cluster Controller Distributed Request Virtual database 2 Authentication Request Request Recovery HSQLDB Recovery Log Scheduler Query result cache Load balancer Scheduler Query result cache Load balancer Recovery Log Recovery HSQLDB PostgreSQL JDBC PostgreSQL JDBC PostgreSQL JDBC PostgreSQL JDBC Postgre SQL Postgre SQL Postgre SQL Postgre SQL The distributed request manager wraps the regular (centralized) virtual database by sending queries in the same total order at all controllers. Queries are then execute locally as if we were in a centralized case. Upon completion, an acknowledgement message is sent to the initiator of the request. 35

36 Sequoia failure handling (1/2) Failure of a connection to a controller transparently reconnected to another controller (according to user defined policy) failure inside a transaction restores the transactional context upon reconnection backend failure backend removed from load balancer (no noticeable failover time) failure during read: retry on another backend failure during write: ignored if successful on other backends 36

37 Sequoia failure handling (2/2) controller failure database backends attached to dead controller are lost clients are transparently reconnected to another controller (according to user defined policy) failure during read: transparently retried (even within a transaction) failure during write: exception thrown to the application telling that the write might have been executed but we cannot be sure 37

38 Node failure with Sequoia execute INSERT INTO t Client application (Servlet, EJB,...) C-JDBC Client application (Servlet, EJB,...) C-JDBC No 2 phase-commit parallel transactions failed nodes are automatically disabled C-JDBC Controller Virtual database 1 Authentication Recovery Log Request Scheduler Query result cache Load balancer JDBC JDBC JDBC If a node fails to execute an update query but other backends succeed, the failing backend is automatically disabled. 38

39 Outline RAIDb Sequoia Geronimo/Sequoia/ Scalability High availability 39

40 Highly available web site Apache clustering L4 switch, RR-DNS, One-IP techniques, LVS, Tomcat clustering mod_jk (T4), mod_proxy/mod_rewrite (T5), session replication clustering C-JDBC RR-DNS Internet mod-jk Parsing cache Result cache Metadata cache embedded Even with a single database backend, C-JDBC can significantly improve performance with its various caches that prevent requests from going to the database. This is well-suite for ecommerce applications that have read-mostly workloads. 40

41 Result cache Cache contains a list of SQL ResultSet Policy defined by querypattern Policy 3 policies EagerCaching: variable granularities for invalidations RelaxedCaching: invalidations based on timeout NoCaching: never cached RUBiS bidding mix No cache Coherent Relaxed with 450 clients cache cache Throughput (rq/min) Avg response time 801 ms 284 ms 134 ms CPU load 100% 85% 20% C-JDBC CPU load - 15% 7% Eager caching accepts database, table, column and column unique granularities. Basically, every request is parsed and C-JDBC checks which tables or rows are needed to compute the result. When a write queries is executed, it invalidates all results that required the data that has been invalidated. The finer grain invalidation requires more extensive parsing (more cpu) but have better hit ratio. Relaxed caching can further enhance the cache performance by relaxing the consistency of queries that can accept a given staleness (for example a page whose content is only refreshed every minute). 41

42 Highly available web applications choosing RAIDb level recovery log for adding nodes dynamically recovering from failures Internet Sequoia 42

43 <connector <resourceadapter> <dependency> configid="my" sequoia/jars/sequoia-.jar Geronimo <name>postgresqldatasource</name> xmlns=" <uri>sequoia/jars/sequoia sequoia/jars/sequoia-.jar parentid="org/apache/geronimo/server">.jar</uri> </dependency> version="1.5" configuration Copy sequoia-.jar in setting><config-property-setting org.continuent.sequoia..driver</config-property-setting> name="password"> name="driver"> name="username"> dbpw</config-property-setting> dbuser</config-property- </connector> jdbc:sequoia://c1,c2/mydb</config-property-setting> </resourceadapter> <config-property-setting name="url"> 43

44 Configuring Sequoia with Network server Virtual database configuration file <Virtual name="mydb"> <Distribution> <MessageTimeouts/> </Distribution> < name= derby1 = org.apache.derby.jdbc.clientdriver url= jdbc:derby:net://localhost:1527/xpetstore;create=tru e;retrievemessagesfromserverongetmessage=true; connectionteststatement="values 1"> < /> 44

45 Configuring Sequoia Clustering with (1/2) <Request> <RequestScheduler> <RAIDb-1Scheduler level= passthrough"/> </RequestScheduler> <RequestCache> <MetadataCache/> <ParsingCache/> <ResultCache granularity="table" /> </RequestCache> <LoadBalancer> <RAIDb-1> <RAIDb-1-LeastPendingRequestFirst/> </RAIDb-1> </LoadBalancer> RAIDb-1 scheduler and load balancer ensures proper database replication. 45

46 Configuring Sequoia Clustering with (2/2) <RecoveryLog> <JDBCRecoveryLog ="org.apache.derby.jdbc.clientdriver url="jdbc:derby:net://localhost:1529/xpetstore; create=true;retrievemessagesfromserverongetmessage=true;" login="app" password="app"> <RecoveryLogTable tablename="recovery" idcolumntype="bigint NOT NULL" sqlcolumnname="sqlstmt sqlcolumntype="varchar(8192) NOT NULL extrastatementdefinition=",primary KEY (id)"/> <CheckpointTable tablename="checkpoint"/> <Table tablename="backendtable"/> <DumpTable tablename= DUMPTABLE /> </JDBCRecoveryLog> </RecoveryLog> </Request> </Virtual> Note that it is also possible to use an embedded for the recovery log. 46

47 Outline RAIDb C-JDBC Geronimo/Sequoia/ Scalability High availability 47

48 Sequoia vertical scalability allows nested RAIDb levels allows tree architecture for scalable write broadcast necessary with large number of backends Sequoia re-injected in Sequoia controller 48

49 Advanced Configurations Vertical Scalability Mixed Horizontal and Vertical Scalability Client program Sequoia JVM Client program Sequoia JVM Client program Sequoia JVM Client program Sequoia JVM Client program Sequoia JVM Client program Sequoia JVM Sequoia controller Partial replication Sequoia Sequoia controller Full replication DB native Sequoia JDBC Sequoia controller Full replication Sequoia Sequoia controller Full replication DB native JDBC Sequoia controller Full replication DB native JDBC Sequoia controller Full replication DB native JDBC DB 1 DB 2 Sequoia controller Full replication DB native Sequoia JDBC DB 1 DB 2 DB 3 DB 4 DB 5 DB 6 DB7 DB 3 DB 4 Sequoia controller Full replication DB native JDBC DB 5 DB 6 DB 7 49

50 TPC-W benchmark (Amazon.com) Nearly linear speedups with the shopping mix Throughput in requests per minute Number of nodes Single Full replication Partial replication TPC-W is a database benchmark that models an online bookstore like amazon.com. More information can be found at 50

51 Outline RAIDb C-JDBC Geronimo/Sequoia/ Scalability High availability 51

52 Building initial checkpoint Dump initial database using any tools (tar, zip, ) Initial checkpoint inserted in RecoveryLog EJB Container JOnAS, WebLogic, JBoss, WebSphere,... C-JDBC JVM disabled dump for initial checkpoint C-JDBC Controller Recovery Log JDBC disabled 52

53 Logging is enabled All database updates are logged (SQL statement, user, transaction, ) JDBC Recovery Log dump for initial checkpoint EJB Container JOnAS, WebLogic, JBoss, WebSphere,... C-JDBC JVM C-JDBC Controller Recovery Log JDBC enabled enabled 53

54 Adding new backends 1/3 Add new backends while system online Restore dump corresponding to initial checkpoint JDBC Recovery Log dump for initial checkpoint EJB Container JOnAS, WebLogic, JBoss, WebSphere,... C-JDBC JVM enabled C-JDBC Controller Recovery Log JDBC disabled enabled disabled 54

55 Adding new backends 2/3 Replay updates from the log EJB Container JOnAS, WebLogic, JBoss, WebSphere,... C-JDBC JVM enabled JDBC Recovery Log dump for initial checkpoint C-JDBC Controller Recovery Log JDBC disabled enabled disabled 55

56 Adding new backends 3/3 Enable backends when done EJB Container JOnAS, WebLogic, JBoss, WebSphere,... C-JDBC JVM enabled JDBC Recovery Log dump for initial checkpoint C-JDBC Controller Recovery Log JDBC enabled enabled enabled 56

57 Making new checkpoints (1/3) Disable one backend to have a coherent snapshot Mark the new checkpoint entry in the log Dump with tar/zip dump for initial checkpoint... JDBC Recovery Log dump for dump last checkpoint for last checkpoint EJB Container JOnAS, WebLogic, JBoss, WebSphere,... C-JDBC JVM enabled C-JDBC Controller Recovery Log JDBC disabled enabled enabled 57

58 Making new checkpoints (2/3) Replay missing updates from log EJB Container JOnAS, WebLogic, JBoss, WebSphere,... C-JDBC JVM enabled dump for initial checkpoint... JDBC Recovery Log dump for dump last checkpoint for for last checkpoint C-JDBC Controller Recovery Log JDBC disabled enabled enabled 58

59 Making new checkpoints (3/3) Re-enable backend when done dump for initial checkpoint... JDBC Recovery Log dump for dump last checkpoint for for last checkpoint EJB Container JOnAS, WebLogic, JBoss, WebSphere,... C-JDBC JVM enabled C-JDBC Controller Recovery Log JDBC enabled enabled enabled 59

60 Handling failures A node fails! Automatically disabled but administrator fix needed dump for initial checkpoint... JDBC Recovery Log dump for dump last checkpoint for for last checkpoint EJB Container JOnAS, WebLogic, JBoss, WebSphere,... C-JDBC JVM enabled C-JDBC Controller Recovery Log JDBC disabled enabled enabled 60

61 Recovery 1/3 Restore latest dump EJB Container JOnAS, WebLogic, JBoss, WebSphere,... C-JDBC JVM enabled dump for initial checkpoint... JDBC Recovery Log dump for dump last checkpoint for last checkpoint C-JDBC Controller Recovery Log JDBC disabled enabled enabled 61

62 Recovery 2/3 Replay missing updates from log dump for initial checkpoint... JDBC Recovery Log dump for dump last checkpoint for for last checkpoint EJB Container JOnAS, WebLogic, JBoss, WebSphere,... C-JDBC JVM enabled C-JDBC Controller Recovery Log JDBC disabled enabled enabled 62

63 Recovery 3/3 Re-enable backend when done dump for initial checkpoint... JDBC Recovery Log dump for dump last checkpoint for for last checkpoint EJB Container JOnAS, WebLogic, JBoss, WebSphere,... C-JDBC JVM enabled C-JDBC Controller Recovery Log JDBC enabled enabled enabled 63

64 Current limitations Distributed joins Out parameters for stored procedures Some JDBC 3.0 extensions XA support through XAPool only network partition/reconciliation not supported 64

65 Conclusion RAIDb RAID-like scheme for databases Sequoia open source middleware for database replication performance scalability high availability Geronimo/Sequoia/ End-to-end open source highly available web applications 65

66 Q&A Thanks to all users and contributors