Yogesh Simmhan. escience Group Microsoft Research

Size: px

Start display at page:

Download "Yogesh Simmhan. escience Group Microsoft Research"

Jade Craig
6 years ago
Views:

1 External Research Yogesh Simmhan Group Microsoft Research Catharine van Ingen, Roger Barga, Microsoft Research Alex Szalay, Johns Hopkins University Jim Heasley, University of Hawaii

2 Science is producing vast quantities of data Data needs to be ingested & published for community use Different user roles, different domain/it expertise Provenance, reliability, semantics are key! Commodity hardware Produc er Raw Data Valet Verified Data Curator Ingested Data Publish er Published Data Consu mer

3 Scientific workflows are popular with scientists Composable & reusable pipelines supporting data flows Diverse execution environments, tracking & monitoring Workflows are suitable for data ingest But reliability is important for valet workflows Workflow design models for reliability in distributed environment Workflow framework features for fault tolerance

Panoramic Survey Telescope And Rapid Response System Discover dangerous asteroids & comets.

month Large, growing data collection Annual 1PB of raw images 30TB of community data 5.

5TB image data 100GB of shared data 1000 images 150M detections The PS1 Telescope at Haleakala, Maui, Hawaii Raw

4 Panoramic Survey Telescope And Rapid Response System Discover dangerous asteroids & comets. Use data for Solar System, cosmology study One of the largest visible light telescopes 2/ 3 rds of sky scanned thrice a month Large, growing data collection Annual 1PB of raw images 30TB of community data 5.5B objects 350B detections Daily 2.5TB image data 100GB of shared data 1000 images 150M detections The PS1 Telescope at Haleakala, Maui, Hawaii Raw Digital Images [~1PB/yr] Image Processing Pipeline (IPP) Unix Cluster Detections, Objects, Metadata (Comma Separated Value files) [~100TB/yr] Object Data Manager (ODM) MS-SQL DB on Win2008 Servers Query Results SQL Query CASJobs Batch Submission Portal ~300 Astronomers

5 Data partitioned as distributed DB with views WinHPC Cluster,MSSQL 3 copies: hot/cold/warm CSV files arrive from Image Pipeline: 1k daily Each file loaded into one Load database Load DBs from each week merged with prior data in Cold database Copies of new Cold DB surfaced to users Workflows for data ops: Cold DB Hot DB S1 S2 S3 S15 S16 L192 RAID Storage Cold DB spatial partitions Database is Logical Store Warm DB S1 S2 S16 S14 S7 S8 S9 S11 S15 S16 S3 S1 CSV Batch CSV Batch Load Load Workflow (900/night) Loads a CSV Batch into a new Load DB (~100MB) Online Servers Load DB Load DB Cold DB Hot/Warm DB spatial partitions Merge Workflow (16/week) Merge Load DBs into a Cold DB (~4TB) Merge Load, Merge, Copy, Flip running on Trident Offline Servers Cold Merged DB Copy Workflow (32/week) N/W copy merged Cold DB to replace Hot/Warm DB (~4TB) Copy Copy Warm DB Warm DB Hot DB Hot DB L581 csv csv csv Flip Flip CSV Batch data product on file system logical store Load DB partition Flip Workflow (2/week) Pause CASJobs & recreate Distributed Partion View over the 16 Hot/Warm DBs Warm DPV Hot DPV

6 Well defined, Well tested workflows Valet workflows run repeatedly, their impact is cumulative Workflows and activities do not change often Changes need to be synchronized with repository schedule Testing is crucial E.g. Load workflow run 900 times/day. Faults can compound. Granular, Reusable workflows Avoid custom activities for ad hoc tasks Easier to test and maintain library of core activities Separate policy from mechanism E.g. static n/w files E.g. Copy & Flip are separate workflows. Preamble for Copy is separate from actual copy action.

7 Workflows as Data State Machines State of repository depends on state of its constituent resources Workflows operate on state machines Data containers have states Workflows and activities cause state transitions Instantly determine state of system Easier to interpret workflow progress, complex interactions Impact of workflows on states are different. E.g. Load vs. Merge workflow failures Easily define fault paths based on data state Goal: Recovery from fault state to clean state

8 Workflows as Data State Machines Simple state definitions depending on stage of loading data, recovery model used Activities should cause at most one state change States used for recovery workflow, policy workflow & display Load Workflow Activities State Machine for Load DB State transitions take place when activities execute. CSV Batch Create DB Success Does Not Exist Drop DB (if exists) Drop DB Success Create DB Failed Created Create DB Insert CSV Success Insert CSV Failed Loading Insert CSV Batch Validate Success Drop DB Failed Validate Failed Validate Data Load Complete Load DB Continues in the Merge workflow Activity Type Blue/Solid : Mainline Code Purple/Dashed: Recovery Preamble Success States Green/Solid : Clean Yellow/Dashed: In Flight Failed States Orange/Dotted: Recoverable Red/Dotted: Unrecoverable

9 Workflow Recovery Baked into Design Faults are a fact of life in distributed systems: hardware loss, transient failure, data errors Recovery as part of normal workflow makes handling faults a routine action Coarser (simpler) approach to recovery Different recovery designs Re-Execute Idempotent Resume Idempotent Recover-Resume Idempotent Independent Recovery

10 Workflow Recovery Baked into Design Re-Execute Idempotent Recovery Idempotent workflows that can be rerun without side-effects Input data states valid, no in-place update Retry limits E.g. Flip workflow for DPV reconstruction Resume Idempotent Recovery Idempotent activities allow goto at start Cost, complexity of resume vs. re-execute E.g. Copy workflow for parallel file copies

11 Workflow Recovery Baked into Design Recover & Resume Separate activity(s) to rollback to initial state Reduce the problem to resume/re-execute Passive Recovery: Rollback activity(s) at start of workflow. E.g. Load workflow drop database Active Recovery: Rollback workflow captures operations to undo dirty state. E.g. Merge workflow on machine failure Fail fast vs. Fault tolerant activities Independent Recovery Complex faults requiring global synchronization Sanity check, cleanup, consistent states Manually coordinated. E.g. Loss of merged Cold DB

12 Specific requirements to support data valet tasks. E.g. tracking more imortant than ad hoc composition Provenance Collection Track operations across data, activities, workflows, distributed temporally & spatially Record inputs and outputs to activities Valet workflows need fine grained, system level provenance (logging) Reliable, scalable provenance collection

13 Provenance Collection for Data State Tracking Provenance provides record of state transitions Provenance can be mined for current state of data resources. Recovery based on it. Direct State Recording: Activities expose the state transition they effect as in/outputs Pro: Easily queried; Con: Reusability, local view Indirect State Recording: External mapping from activity execution to state transition Pro: Reusable, WF level; Con: Complexity PS uses indirect model

14 Provenance Collection for Forensics Investigate cause of distributed, repetitive faults Low level, system provenance ~monitoring, logging, tracing: disk, machine, services, I/O Scalable & Resilient Provenance Collection Long running, continuous workflow execution Fine-grained collection Stored till data release, even instrument lifetime. E.g. Status of 30,000 Load workflows Efficient querying for recovery, current state Provenance loss Doubtful repository state PS has realtime SQL replication of provenance

15 Support for Workflow Recovery Models Re-run previous workflow with same inputs Well defined exceptions & recovery execution paths E.g Recover & Resume workflows Identify exception source. Transmit it, reliably, to initiate recovery. Low level, system provenance ~monitoring, logging, tracing: disk, machine, services, I/O Fail-fast guarantees Workflows, activities must fail fast Framework must halt on errors: timeout, liveliness, network connectivity Early, accurate detection Early recovery Trident Workflow configured to fail fast for PS Synchronous provenance logging to Registry HPC fault event monitoring

16 Reliability of repository data management less studied compared to other aspects Data reliability Replication in Grids, Clouds replicated file system, database Transparency vs. Update, recovery overhead Specialized for repositories operate only by workflows, not general purpose applications Application reliability Over-provisioning workflows. Infeasible. Checkpoint restart. Insufficient. Transactional workflows. Do not support data states.

Data management in is complex Large shared data respositories on commodity clusters more common Data valets have special needs from workflows Goal driven approach to

17 Data management in is complex Large shared data respositories on commodity clusters more common Data valets have special needs from workflows Goal driven approach to workflow design using data state machines Simple models for reliability & recovery go a long way Trident workflow provides tools to support both scientist & valet users

18 Acknowledgements Maria Nieto-Santisteban, Richard Wilton, Sue Werner, Johns Hopkins University Conrad Holmberg, University of Hawaii Dean Guo, Nelson Araujo, Jared Jackson, Microsoft Research Trident Workflow Team External Research

Pan-STARRS Introduction to the Published Science Products Subsystem (PSPS)

Pan-STARRS Introduction to the Published Science Products Subsystem (PSPS) The Role of PSPS Within Pan-STARRS Space... is big. Really big. You just won't believe how vastly hugely mindbogglingly big it