The Design of the Borealis Stream Processing Engine CIDR 2005 Brandeis University, Brown University, MIT Kang, Seungwoo Ref. Borealis.ppt

Outline Motivation and Goal Requirements Technical issues  Dynamic revision of query results  Dynamic query modification  Flexible and scalable optimization  Fault tolerance Conclusion

Motivation and Goal Envision the second-generation SPE  Fundamental requirements of many streaming applications not supported by first-generation SPE  Three fundamental requirements Dynamic revision of query results Dynamic query modification Flexible and scalable optimization Present the functionality and preliminary design of Borealis

Requirements Dynamic revision of query results Dynamic query modification Flexible and highly scalable optimization

Changes in the models Data model  Revision support Three types of messages Query model  Support modification of Op box semantic Control lines QoS model  Introduce QoS metrics in tuple basis Each tuple (message) holds VM (Vector of Metrics) Score function

Architecture of Borealis Modify and extend both systems, Aurora and Medusa, with a set of features and mechanisms

1.Dynamic revision of query results Problem  Corrections of previously reported data from data source  Highly volatile and unpredictable characteristics of data sources like sensors Late arrival, out-of-order data  Just accept approximate or imperfect results Goal  Support to process revisions and correct the previous output result

Causes for Tuple Revisions Data sources revise input streams: “On occasion, data feeds put out a faulty price [...] and send a correction within a few hours” [MarketBrowser] Temporary overloads cause tuple drops Data arrives late, and miss its processing wnd Average price 1 hour (1pm,$10) (2pm,$12)(3pm,$11) (4:05,$11)(4:15,$10)(2:25,$9)

New Data Model for Revisions time : tuple timestamp type : tuple type  insertion, deletion, replacement id : unique identifier of tuple on stream headerdata (time,type,id,a1,...,an)

Revision Processing in Borealis Connection points (CPs) store history (diagram history with history bound) Operators pull the history they need  Operator CP Oldest tuple in history Most recent tuple in history Revision Input queue Stream

Discussion Runtime overhead  Assumption Input revision msg < 1% Revision msg processing can be deferred  Processing cost  Storage cost Application’s needs / requirements  Whether it is interested in revision msg or not  What to do with revision msg Rollback, compensation ??  Application’s QoS requirement  Delay bound

2. Dynamic query modification Problem  Change certain attributes of the query at runtime  Manual query substitutions – high overhead, slow to take effect Goal  Low overhead, fast, and automatic modifications Dynamic query modification  Control lines Provided to each operator box Carry messages with revised box parameters and new box functions  Timing problem Specify precisely what data is to be processed according to what control parameters Data is ready for processing too late or too early  old data vs. new parameter -> buffered old parameters  new data vs. old parameter -> revision message and time travel

Time travel Motivation  Rewind history and then repeat it  Move forward into the future Connection Point (CP) CP View  Independent view of CP  View range Start_time Max_time  Operations Undo: deletion message to rollback the state to time t Replay  Prediction function for future travel

3.Optimization in a Distributed SPE Goal: Optimized resource allocation Challenges:  Wide variation in resources High-end servers vs. tiny sensors  Multiple resources involved CPU, memory, I/O, bandwidth, power  Dynamic environment Changing input load and resource availability  Scalability Query network size, number of nodes

Quality of Service A mechanism to drive resource allocation Aurora model  QoS functions at query end-points  Problem: need to infer QoS at upstream nodes An alternative model  Vector of Metrics (VM) carried in tuples Content: tuple’s importance, or its age Performance: arrival time, total resource consumed,..  Operators can change VM  A Score Function to rank tuples based on VM  Optimizers can keep and use statistics on VM

Example Application: Warfighter Physiologic Status Monitoring (WPSM) Area State Confidence Thermal Hydration Cognitive Life Signs Wound Detection 90% 60% 100% 90% 80% Physiologic Models   

SF(VM) = VM.confidence x ADF(VM.age) age decay function Score Function Sensors Model1 Model2 ([age, confidence], value) VM Merge Models may change confidence. HRate RRate Temp

Borealis Optimizer Hierarchy End-point Monitor(s)Global Optimizer Local Monitor Local Optimizer Neighborhood Optimizer Borealis Node statisticstriggerdecision

Optimization Tactics Priority scheduling - Local Modification of query plans - Local  Changing order of commuting operators  Using alternate operator implementations Allocation of query fragments to nodes - Neighborhood, Global Load shedding - Local, Neighborhood, Global

Priority scheduling Highest QoS Gradient box  Evaluate the predicted-QoS score function on each message with values in VM  average QoS Gradient for each box By comparing average QoS-impact scores between the inputs and outputs of each box Highest QoS-impact message from the input queue  Out-of-order processing  Inherent load shedding behavior

Correlation-based Load Distribution Under the situation that network BW is abundant and network transfer delays are negligible Goal: Minimize end-to-end latency Key ideas:  Balance load across nodes to avoid overload  Group boxes with small load correlation together  Maximize load correlation among nodes Connected Plan AB S1 CD S2 r 2r 2cr 4cr Cut Plan S1 S2 AB CD r 2r 3cr

Load Shedding Goal: Remove excess load at all nodes and links Shedding at node A relieves its descendants Distributed load shedding  Neighbors exchange load statistics  Parent nodes shed load on behalf of children  Uniform treatment of CPU and bandwidth problem Load balancing or Load shedding?

Local Load Shedding PlanLoss LocalApp 1 : 25% App 2 : 58% CPU Load = 5Cr 1, Cap = 4Cr 1 CPU Load = 4Cr 2, Cap = 2Cr 2 A must shed 20%B must shed 50% 4CC C 3C r1r1 r1r1 r2r2 r2r2 App 1 App 2 Node ANode B Goal: Minimize total loss 25% 58% CPU Load = 3.75Cr 2, Cap = 2Cr 2 B must shed 47% No overload A must shed 20% r2’r2’

Distributed Load Shedding CPU Load = 5Cr 1, Cap = 4Cr 1 CPU Load = 4Cr 2, Cap = 2Cr 2 A must shed 20%B must shed 50% 4CC C 3C r1r1 r1r1 r2r2 r2r2 App 1 App 2 Node ANode B 9% 64% PlanLoss LocalApp 1 : 25% App 2 : 58% DistributedApp 1 : 9% App 2 : 64% smaller total loss! No overload A must shed 20% r2’r2’ r 2 ’’ Goal: Minimize total loss

Fault-Tolerance through Replication Goal: Tolerate node and network failures U  Node 1’  Node 2’ Node 3’  U s1s1 s2s2 Node 1 U  Node 3 U  Node 2  s3s3 s4s4 s3’s3’

Fault-Tolerance Approach If an input stream fails, find another replica No replica available, produce tentative tuples Correct tentative results after failures STABLE UPSTREAM FAILURE STABILIZING Missing or tentative inputs Failure heals Another upstream failure in progress Reconcile state Corrected output

Conclusions Next generation streaming applications require a flexible processing model  Distributed operation  Dynamic result and query modification  Dynamic and scalable optimization  Server and sensor network integration  Tolerance to node and network failures