1

Crash-Only Web Services:

Failure Semantics in an SOA Environment

www.oasis-open.org

Chris Hobbs and Abbie Barbir

Nortel OASIS Symposium 2007, San Diego

2

The crash-only model Software design

approach Easier to restart

quickly in a known state than to clean up and rebuild to recover from an error

George Candea and Armando Fox arekey proponents of crash-only software

3

Two themes of this talk Discuss issues of the behaviors of individual

and composed services and their part in Web Services Service Level Agreements (WSLA)

Based on the behaviors of the individual services Need a taxonomy or ontology of service behaviors Need an approach to calculating behaviors of

composed services The “crash-only” model of operation as a

simple failure behavior for a Web Service Failure is one of many identified behaviors

4

Background: Orchestration as a New Programming Paradigm

SOA promotes the concept of combining services through orchestration - invoking services in a defined sequence to implement a business process

Orchestration compounds the difficulties of testing and managing the quality of the deployed services

Testing composite services in SOA environment is a discipline which is still at an early stage of study

Describing and usefully modeling the individual and combined behaviors - needed to offer Service Level Agreements (SLA) - is at an even earlier stage

We hope to stimulate additional research on these topics

5

Testing Composed Services It’s fairly straightforward to test the operation of

a device or system if we control all the parts. When we start offering orchestrated services

as a product, the services we are using may be outside our control.

For example consider well-known components: Google mapping service Amazon S3 storage service Mobile operator’s location service

6

Testing Composed Services (2) With orchestrated services, there is never a

complete “box” we can test With orchestration as the new programming

paradigm, testing becomes a much bigger problem

Failures of orchestrated services are often “Heisenbugs” - impervious to conventional debugging, generally non-reproducible

Offering a WSLA based on testing alone, without reliable knowledge of component service behaviors, may be risky

7

Web Services SLA (WSLA)

Concerned with behaviors of the message flows and services spanning the end-to-end business transaction

Clients can develop testing strategies that stress the service to ensure that the service provider has met the contracted WSLA commitment

Composed services make offering a WSLA more risky

Service Provider Z

Provider XService X

Provider YService Y

Client

WSLA

Network

Packets

Message flows

Web Service

8

How can WSLAs be derived from behaviors of component services?

Need to develop a model of the behavioral attributes of the individual component Web Services which contribute to the overall behavior of an orchestrated or composed Web Service.

Need to model the combination of individual service behavioral models

9

Web Services behaviors Behaviors may be described

and quantified for each Web Service

May be combined by a “calculus of behaviors” when multiple services are composed

Behavior parameters may become a part of the service description, perhaps in WSDL.

Availability and Reliability

Performance Management Failure Security Privacy, confidentiality

and integrity Scalability Execution Internationalization Synchronization Etc., …

10

Web Services behaviors (2) To develop a Service Level

Agreement (SLA) for a composed service (Z), we need to have relevant behavior descriptions for the individual services (X and Y)

We also need a deep understanding of how to combine the descriptions of X and Y to calculate results for Z

ZX Y

11

Web Services behaviors (3) For each behavior, the challenges include the

following: 1. How may service X’s and service Y’s behavior

be characterized? 2. How may those characterizations be formalized

and advertised by X and Y? 3. How may Z incorporate X’s and Y’s

characterizations and then advertise the result?

Z itself might become a component of an even larger service and therefore needs to advertise its own characteristics. It also needs this characterization to offer an SLA to consumers.

12

Web Services behaviors (4) Each behavior may have its own ontology,

measures, and calculus of combining those measures when services are composed.

Local Ontology

Local Ontology

Local Ontology

Abstracted Ontology

Abstracted Ontology

X

Y

Z

Z – Specific Ontology

?Need this analysis for each behavior

of services X, Y and Z

13

Web Services behaviors (5) Ten behavior examples

Availability and Reliability Performance Management Failure (Crash-only is one mode) Security Privacy, confidentiality and integrity Scalability Execution Internationalization Synchronization

Let’s focus on a few of these behaviors…

Source: “Advertising Service Properties,” unpublished paper by C. Hobbs, J. Bell, P. Sanchez

14

Availability and Reliability “Availability” is the percentage of client

requests to which the server responds within the time it advertised.

“Reliability” is the percentage of such server responses which return the correct answer.

In some applications availability is more important than reliability

Many protocols used within the Internet, for example, are self-correcting and an occasional wrong answer is unimportant. The failure to give any answer, however, can cause a major network upheaval.

15

Availability and Reliability (2) In other applications reliability is

more important than availability If the service which calculates a

person’s annual tax return does not respond occasionally it’s not a major problem - the user can try again

If that service does respond but with the wrong answer which is submitted to the tax authorities, then it could be disastrous

16

Availability and Reliability (3)

Services are built with either availability or reliability in mind, with clients accepting that no service can ever be 100% available or 100% reliable.

In combining services X and Y into a composite service Z, it is necessary to combine the underlying availability and reliability models and predict Z’s model.

To do so without manual intervention, X’s and Y’s models must be exposed.

17

Availability and Reliability (4) Availability and reliability models are

often expressed as Markov Models or Petri Nets, which are easy to combine in a hierarchical way.

Major issues: Agreeing upon the semantics of the states

in the Markov model or places in the Petri nets

Finding a way for X and Y to publish the models in a standard form.

18

Availability and Reliability (5) Currently, apart from raw percentage figures,

there is no method for describing these models

Percentage time when the server is unavailable? Percentage of requests to which it does not

reply? Different clients may experience these differently A server which is unavailable from 00:00 to 04:00

every day can be 100% available to a client that only tries to access it in the afternoons.

19

Availability and Reliability (6)

If X and Y are distributed, then it is possible, following network failures, that for some customers, Z can access X but not Y and for others Y but not X.

The assessment of Z’s availability may be hard to quantify, so it may be difficult for Z to offer a meaningful WSLA.

20

Failure

The failure models of X and Y may be very different:

X fails cleanly and may, because of its idempotency, immediately be called again

Y has more complex failure modes Z will add its own failure modes to those of X and Y Predicting the outcome could be very difficult

The complexity is increased because many developers do not understand failure modeling and, even were models to be published, their combination would be difficult due to their stochastic nature.

21

Failure (2) One approach to describing a service’s failure

model: Service publishes the exceptions that it can raise

and associates the required consumer behavior with each

“Exception D may be thrown when the database is locked by another process. Required action is to try again after a random backoff period of not less than 34ms.”

“Crash-only” failure model is a simple starting point for building a taxonomy of failure behavior. This work is just beginning.

22

Scalability

A behavioral description and WSLA for the composite service Z must include its scalability

How many simultaneous service instances can it support?

What service request rate does it handle? etc. These parameters will almost certainly differ

between the component services X and Y, and will need to be published by those services.

X and Y are presumably not dedicated solely to Z, so the actual load being applied to X and Y at any given time is unknown to the provider of Z, making the scalability of Z even harder to determine.

23

Web Services behaviors (again) Ten behavior examples

Availability and Reliability Performance Management Failure (Crash-only is one mode) Security Privacy, confidentiality and integrity Scalability Execution Internationalization Synchronization

We described a few of these behaviors… Can we use them to build WSLAs?

24

Web Service Level Agreement (WSLA)

Based on behaviors and descriptors for these behaviors.

Example: Failure model Is transaction half-performed? Is it re-wound?

These behaviors and descriptors are not available in the WS description, in WSDL No performance info Not even price!

25

Web Service Level Agreements (2) Business acceptance of composed services for

business-critical operations depends on a service provider’s ability to offer WSLA

Uptime, response time, etc. Offering a WSLA depends on ability to compose the

WSLA-related behaviors of the individual services This information needs to be available via WSDL or

similar source Should include test vectors to test the SLA claims

The ability to determine and offer a WSLA commitment is a limiting factor for widespread acceptance of services based on orchestration

26

Web Service Level Agreements – conclusions

Need a more precise way to express the parameters of behaviors

Availability – What is 99.97% uptime? Several milliseconds outage each minute? Several minutes planned downtime each month?

Failure model – Crash-only as the simplest, lowest layer or level of failure in a future full failure model.

Eight other SLA-related behaviors listed here – each has a complex semantic for description and composition

More questions than answers now - many PhDs still to be earned in this area!

27

Back to the crash-only software model

Can it simplify service composition, testing, development of WSLA, and end world hunger?

28

Crash-only software (1)

Historically, developers have spent a lot of effort making software resilient

Put borders around it so it will not affect other things if it fails

Try to close it down cleanly Save state Reload the software component Restart and replay

Trying to keep the client from becoming aware that a failure occurred

29

Crash-only software (2) Years of work over last ten years on

resilient software - which stays up all the time, and recovers from problems For example, tutorials by Bev Littlewood

Crash-only software is the exact opposite Client accepts that the server may crash

Power failure, network down, hardware, etc. Client must be able to recover or restart the

process by itself

30

Crash-only software (3) Crash-only principles

Forget recovery - more trouble than it’s worth When the server senses a problem, it will “crash” as

cleanly as possible and may perform a “micro-reboot” to return to original state

Sometimes recover to a well-defined checkpoint Client may initiate the crash

The server is back working sooner than if it tried to recover via logs and journals, etc.

Principles fit the Web Services paradigm nicely! Loose coupling of services Little state shared among services

31

Crash-Only Software (4) Crash-only semantic has several advantages:

Simpler macroscopic behavior with fewer externally visible states

Reduces outage time by removing all shutting-down time Simplifies failure model by reducing recovery state table

size Crashing can be invoked from outside the software of

the provider Recovery from a failed state is notoriously difficult and

the crash-only paradigm coerces the system into a known state without attempting recovery

Reduce the complexity of the provider code Simplifies testing by reducing the failure combinations

that have to be verified. Consumer is assumed to be able to initiate the crash.

32

Crash-Only Web Services

Candea’s list of properties required for a crash-only system can be abstracted to match properties of Web Services:

Components have externally enforced boundaries. This is supported by the virtual machine concept used on many Web Service systems

All interactions between components have a timeout. This is implicit in any loosely-coupled Web Services interaction.

All resources are leased to the service rather than being permanently allocated. This is particularly useful in Web Services.

Requests are entirely self-describing. For crash-only services this requires that the request carries information about time-to-live and idempotency – will it return the same result if invoked again?.

All important non-volatile state is managed by dedicated state stores.

33

Crash-Only Reliable Web Service

For systems with hardware redundancy, by using crash only techniques, SOAP & WS-RM can be extended in order to produce an always available Web Service from the provider’s and consumer’s point of view

WSLA response time may be at risk if a service is forced to crash

Crash-Only Application

Server

Stall Proxy

Web ServiceConsumer

Web ServicesEndpoint

RecoveryAgent

Crash-Only Backend

Crash-Only Backend

Crash-onlyWSM

Internet

Internet

Reliable SOAP Protocol

WS-ReliableMessaging

34

Conclusions Testing Web Services in an SOA environment is

a discipline that is still in its infancy There are no standard models to describe or

combine Web Services behavior information across various services and providers

Web Services SLAs (WSLAs) for composed services are problematic

Testing is only a partial solution Behavioral composition needs work, but is promising

Crash-only Web Services can address some of these difficulties

There are many related areas for further work

35

Q & AQ & A