EECS 498 Introduction to Distributed Systems

Transcription

1 EECS 498 Introduction to Distributed Systems Fall 2017 Harsha V. Madhyastha

2 November 27, 2017 EECS 498 Lecture 19 2

3 Windows Azure Storage Focus on storage within a data center Goals: Durability Strong consistency High availability Scalability Primary-backup replication November 27, 2017 EECS 498 Lecture 19 3

4 Revisiting Project 2 Assume viewservice is Paxos-based RSM Why would client s Append fail? Viewservice may not have detected failure Key idea in Azure: Client triggers view change Why didn t we do this in Project 2? View change was a heavyweight operation November 27, 2017 EECS 498 Lecture 19 4

5 Thought experiment What if we supported only Append and Get operations in Project 2? How to leverage this to make view change lightweight? Hint: All replicas don t fail when view change is initiated Don t transfer any state upon view change Serve GETs based on state from all views November 27, 2017 EECS 498 Lecture 19 5

6 WAS: High-level idea Treat state as append-only log PUTs from clients applied as appends to log Log comprises concatenation of extents Only last extent can be modified;; all others read-only View change has two steps: Seal last extent Add new extent November 27, 2017 EECS 498 Lecture 19 6

7 Logical view of data in WAS Stream //foo/myfile.dat Ptr E1 Ptr E2 Ptr E3 Ptr E4 Ptr E5 Extent E1 Extent E2 Extent E3 Extent E4 Extent E5 November 27, 2017 EECS 498 Lecture 19 7

8 Creating an Extent Client Create Stream/Extent EN1 Primary EN2, EN3 Secondary SM SMView Service Paxos Allocate Extent replica set EN 1 EN 2 EN 3 EN Primary Secondary A Secondary B November 27, 2017 EECS 498 Lecture 19 8

9 Replication Flow Client EN1 Primary EN2, EN3 Secondary SM SMView Service Paxos Ack Append Upon failure, how to determine size of extent before sealing it? EN 1 EN 2 EN 3 EN Primary Secondary A Secondary B November 27, 2017 EECS 498 Lecture 19 9

10 Extent Sealing (Scenario 1) Seal Extent Client Seal Extent SM Stream SM Master Sealed at 120 Append Ask for current length EN 1 EN 2 EN 3 EN 4 Primary Secondary A Secondary B November 27, 2017 EECS 498 Lecture 19 10

11 Extent Sealing (Scenario 1) Client SM Stream SM Master Seal Extent Sealed at Sync with SM EN 1 EN 2 EN 3 EN 4 Primary Secondary A Secondary B November 27, 2017 EECS 498 Lecture 19 11

12 Extent Sealing (Scenario 2) Client Seal Extent SM SMSM Seal Extent Append Ask for current length EN 1 EN 2 EN 3 EN 4 Primary Secondary A Secondary B Scenario in which servers would differ in responses? November 27, 2017 EECS 498 Lecture 19 12

13 Extent Sealing (Scenario 2) Seal Extent Client Seal Extent SM SMSM Sealed at 100 Append Ask for current length EN 1 EN 2 EN 3 EN 4 Primary Secondary A Secondary B November 27, 2017 EECS 498 Lecture 19 13

14 Extent Sealing (Scenario 2) Client SM SMSM Seal Extent Sealed at Sync with SM EN 1 EN 2 EN 3 EN 4 Primary Secondary A Secondary B November 27, 2017 EECS 498 Lecture 19 14

15 Serving GETs How to lookup latest value of any key? Unlike thought experiment, WAS supports PUTs Ask Stream Master (aka view service)? Stream Master only has definitive copy of metadata: Number of extents, Length of sealed extents, Unaware of what updates clients performed November 27, 2017 EECS 498 Lecture 19 15

16 Serving GETs Need to maintain index of data in each stream which offset in which extent has latest value for key? How to execute a GET on a key? Lookup index for (extent, offset), then lookup stream Implication for execution of PUTs? Must not only append to stream, but also update index November 27, 2017 EECS 498 Lecture 19 16

17 Layering in Azure Storage Access blob storage via the URL: Storage Location Service LB LB Front-Ends Front-Ends Partition Layer Inter-stamp (Geo) replication Partition Layer Stream Layer Intra-stamp replication Storage Stamp Stream Layer Intra-stamp replication Storage Stamp November 27, 2017 EECS 498 Lecture 19 17

18 Partition Layer Index Range Partitioning Blob Index Account Name Container Name Blob Name aaaa aaaa aaaaa Account harry.. Container pictures.. sunrise.. Blob Name Name Name.. Front-End.... harry pictures sunset.. Server A- H: PS1 Account.. H - R: Container.. PS2.. Blob richard Name videos Name soccer Name.. R - Z:.. PS3.. richard videos tennis Partition Map zzzz zzzz zzzz zzzz zzzzz zzzzz Storage Stamp Partition Server A- H PS 1 Partition Server H - R PS 2 A- H: PS1 Partition H - R: PS2 Master R - Z: PS3 Partition Server R - Z Partition Map PS 3 November 27, 2017 EECS 498 Lecture 19 18

19 Overall Architecture Front End Layer Incoming Write Request Ack FE FE FE FE FE Partition Master Lock Service Partition Layer Partition Server Partition Server Partition Server Partition Server M Stream Layer M Paxos M Extent Nodes (EN) November 27, 2017 EECS 498 Lecture 19 19

20 Beating CAP Theorem Stream layer Availability: Seal extent and switch to new replica set Consistency: Replicas agree on commit length Partition layer Consistency: Disjoint partition ranges across servers Availability: Partition servers only cache state Key enabler: Append-only design Drawbacks? Garbage collection overhead November 27, 2017 EECS 498 Lecture 19 20

21 Enabling Scalability Use replicated Paxos-based service to make centralized decisions Stream Master Partition Master Lock service at partition layer But Ensure central service is off the data path Cache state from central service when possible November 27, 2017 EECS 498 Lecture 19 21

22 Replica Selection How should stream master choose 3 replicas for an extent? Different copies in different racks Ideally, racks with different sources of power November 27, 2017 EECS 498 Lecture 19 22

23 Optimizing Read Latency Can read from any replica of a sealed extent What if replica that a client chooses to read from is currently under high load? How to reduce latency impact? Read from all 3 replicas;; wait for first response 3x increase in load Read from any replica;; give deadline to respond Retry at another replica if deadline violation November 27, 2017 EECS 498 Lecture 19 23

24 Optimizing Storage Cost Storing 3 replicas à 3x storage cost How to reduce amount of data stored? Exploit benefit of sealing an extent: Content of sealed extent never changes in future November 27, 2017 EECS 498 Lecture 19 24

25 Erasure Coding + + = Split data into M chunks Use coding to generate N additional chunks Data recoverable using any M of (M + N) chunks How many chunks must be lost to lose data? Better durability than 3x replication at x storage overhead November 27, 2017 EECS 498 Lecture 19 25

26 Combining Optimizations How to address tradeoff? Erasure coding: Lower storage cost Replication: Lower read latency For any extent Store one original copy Store (M + N) erasure coded chunks If read from replica misses deadline, reconstruct data from coded chunks November 27, 2017 EECS 498 Lecture 19 26