Availability in Globally Distributed Storage Systems

-
This paper by Google Research provides a report on the availability behavior of large cloud storage systems by studying up to 7K nodes at Google over a year. The paper does not propose any new protocols, but provides sufficiently accurate models of system behavior/performance based on the study. These models are important since they enable us to correctly design and optimize these multi layered systems for data availability.

The work is divided into two parts. The first is the analysis of the component availability (disks, machines, racks), and the second is the analysis of the data availability, as inferred from the component availability results and design decisions of the distributed storage system.

Preliminaries
A storage node is defined as unavailable when it fails to respond positively to periodic health checking pings sent by the monitoring system. The paper does not investigate about network errors specifically; those are also swept under unavailability with software & hardware faults at the node level. Even in Section 3, when the causes of node availability are analyzed in detail, network errors are not mentioned.

Node unavailability does not directly imply data unavailability thanks to identical replication or alternatively erasure encoding. In both approaches, data is divided into a set of stripes, each of which comprises of fixed size chunks. Data in a stripe can be reconstructed from some subsets of the chunks. For replication, R = n refers to n identical chunks in a stripe, so the data may be recovered from any one chunk. For Reed-Solomon erasure encoding, RS(n,m) denotes n distinct data blocks and m error correcting blocks in each stripe. In this case a stripe may be reconstructed from any n chunks.

Node failures and Correlated failures
The majority of node unavailability is due to planned reboots (such as kernel version upgrades), which is followed closely by node restarts (OS restart on the storage node). Unplanned machine reboots (e.g., kernel crashes) don't occur frequently but have the longest average duration to recover since extra checks and corrections are used to put the machines to a safe state.
The paper then focuses on measuring the frequency and severity of correlated failures. (Root causes of these failures --power outages, rolling reboots/upgrades-- are not investigated in the paper.) A failure burst is defined with respect to a window size as a maximal sequence of node failures, each one occurring within a time window 2 minutes of the next. 2 minutes is derived from a knee analysis of the graph plotting the "effect of window size on the fraction of individual failures that get clustered into bursts".

Next, the paper focuses on detecting rack-related node failures. To this end, rack-affinity score is defined as the "probability that a burst of the same size affecting randomly chosen nodes in that cell will have a smaller burst score". For a random burst, the expected rack affinity score is 0.5, for a rack-correlated burst close to 1, and a rack-anti-correlated burst close to 0 (they have not observed any rack-anti-correlated bursts). The authors found that larger failure bursts have higher rack affinity: All the failure bursts of more than 20 nodes have rack affinity greater than 0.7, and those of more than 40 nodes have affinity at least 0.9. Figure 8 shows frequency of failure bursts sorted by racks and nodes affected, and we can clearly see a large fraction of failures are bursty and rack-correlated. Later at Figure 10, it is shown that large failure bursts are the biggest contributor to unavailability as well.
Coping with failures
When a node failure causes the unavailability of a chunk within a stripe, the system initiates a recovery operation for that chunk from the other available chunks remaining in the stripe. The rate at which missing chunks may be recovered is limited by the bandwidth of individual disks, nodes, and racks, and there is a tradeoff between doing recovery and serving new client requests. Of course, given the frequency of bursty rack-correlated failures, it is no surprise that a rack-aware stripe placement policy (that ensures that no two chunks in a stripe are placed on nodes in the same rack) increases the stripe MTTF (i.e., data availability) by a factor of 3 typically.

Markov model of data availability and its findings
Using the data collected and analyzed above for node failures and rack-correlated bursty failures, the paper formulates a Markov model for data availability and develops analytical models to reason about past and future availability in the storage clusters, including the effects of different choices of replication, data placement and system parameters. A very simple example for identical replication and 2 chunks per stripe (i.e., R=2) is given in Figure 12.
The paper shows that the model is able to capture well the effect of failure bursts on the MTTF. Next the paper investigates how changes in the parameters of the system will affect data availability. Using the model the paper finds that reducing recovery times is effective when correlated failures are few. For RS(6,3) with no correlated failures, a 10% reduction in recovery time results in a 19% reduction in unavailability. However, when correlated failures are taken into account, even a 90% reduction in recovery time results in only a 6% reduction in unavailability. Correlated failures (compared to independent/random failures) are also shown to reduce the MTTF by at least two orders of magnitude. Correlation also reduces the benefit of increasing data redundancy. The gain in availability achieved by increasing the replication number, for example, grows much more slowly when we have correlated failures.

Two other important conclusions from the model, that should guide design decisions in distributed storage systems, are as follows. The model shows that component availability improvements below the node (server) layer of the storage stack do not significantly improve data availability. Assuming R=3 is used, a 10% reduction in the disk failure rate increases stripe availability by less than 1.5%. On the other hand, cutting node failure rates by 10% can increase data availability by 18%. The model also shows that replicating data across multiple cells (data centers) greatly improves availability because it protects against correlated failures. For example, R=2x2 (i.e., replicating twice in two cells) with 1 day recovery time between cells has two orders of magnitude longer MTTF than R=4.

Conclusions
I expect this paper to be very influential for distributed storage systems research. Using logs obtained from a large-scale production environment, the paper provided evidence that correlation among node failures dwarfs all other contributions to unavailability. (That said, I am disappointed that the network-related failures are not investigated separately.) The provided analytical model of data availability is very promising; and should be used by the distributed storage systems researchers to guide and evaluate any protocols they propose.

Comments

Unknown said…
This comment has been removed by a blog administrator.
December 17, 2012 at 3:09 AM
Post a Comment

Popular posts from this blog

Hints for Distributed Systems Design

-
This is with apologies to Butler Lampson, who published the " Hints for computer system design " paper 40 years ago in SOSP'83. I don't claim to match that work of course. I just thought I could draft this post to organize my thinking about designing distributed systems and get feedback from others. I start with the same  disclaimer Lampson gave. These hints are not novel, not foolproof recipes, not laws of design, not precisely formulated, and not always appropriate. They are just hints.  They are context dependent, and some of them may be controversial. That being said, I have seen these hints successfully applied in distributed systems design throughout my 25 years in the field, starting from the theory of distributed systems (98-01), immersing into the practice of wireless sensor networks (01-11), and working on cloud computing systems both in the academia and industry ever since. These heuristic principles have been applied knowingly or unknowingly and has proven
Read more >>

Learning about distributed systems: where to start?

-
This is definitely not a "learn distributed systems in 21 days" post. I recommend a principled, from the foundations-up, studying of distributed systems, which will take a good three months in the first pass, and many more months to build competence after that. If you are practical and coding oriented you may not like my advice much. You may object saying, "Shouldn't I learn distributed systems with coding and hands on? Why can I not get started by deploying a Hadoop cluster, or studying the Raft code." I think that is the wrong way to go about learning distributed systems, because seeing similar code and programming language constructs will make you think this is familiar territory, and will give you a false sense of security. But, nothing can be further from the truth. Distributed systems need radically different software than centralized systems do.  --A. Tannenbaum This quotation is literally the first sentence in my distributed systems syllabus. Inst
Read more >>

Making database systems usable

-
Image
C. J. Date's Sigmod 1983 keynote, "Database Usability", was prescient. Usability is the most important thing to the customers. They care less about impressive benchmarks or clever algorithms, and more about whether they can operate and use a database efficiently to query, update, analyze, and persist their data with minimal headache. (BTW, does anyone have a link to the contents of this Sigmod'83 talk? There is no transcript around, except for this short abstract .) The paper we cover today is from Sigmod 2007. It takes on the database usability problem raised in that 1983 keynote head-on, and calls out that the king is still naked.  Let's give some context for the year 2007. Yes, XML format was still popular then. The use-case in the paper is XQuery. The paper does not contain any  reference to json. MongoDB would be released in 2009 with the document model; and that seems to be great timing for some of the usability pains mentioned in the paper! Web 2.0 was in
Read more >>

Looming Liability Machines (LLMs)

-
As part of our zoom reading group ( wow, 4.5 years old now ), we discussed a paper that uses LLMs for automatic root cause analysis (RCA) for cloud incidents. This was a pretty straightforward application of LLMs. The proposed system employs an LLM to match incoming incidents to incident handlers based on their alert types, predicts the incident's root cause category, and provides an explanatory narrative. The only customization is through prompt-engineering. Since this is a custom domain, I think a more principled and custom-designed  machine learning system would be more appropriate rather than adopting LLMs. Anyways, the use of LLMs for RCAs spooked me vicerally. I couldn't find the exact words during the paper discussion, but I can articulate this better now. Let me explain. RCA is serious business Root cause analysis (RCA) is the process of identifying the underlying causes of a problem/incident, rather than just addressing its symptoms. One RCA heuristic is asking 5 Why&#
Read more >>

Foundational distributed systems papers

-
I talked about the importance of reading foundational papers last week. To followup, here is my compilation of foundational papers in the distributed systems area. (I focused on the core distributed systems area, and did not cover networking, security, distributed ledgers, verification work etc. I even left out distributed transactions, I hope to cover them at a later date.)  I classified the papers by subject, and listed them in chronological order. I also listed expository papers and blog posts at the end of each section. Time and State in Distributed Systems Time, Clocks, and the Ordering of Events in a Distributed System. Leslie Lamport, Commn. of the ACM,  1978. Distributed Snapshots: Determining Global States of a Distributed System. K. Mani Chandy Leslie Lamport, ACM Transactions on Computer Systems, 1985. Virtual Time and Global States of Distributed Systems.  Mattern, F. 1988. Practical uses of synchronized clocks in distributed systems. B. Liskov, 1991. Expository papers
Read more >>

Advice to the young

-
Image
I notice I haven't written any advice posts recently. Here is a collection of my advice posts pre 2020. I've been feeling all this elderly wisdom pent up in me, ready to pour at any moment. So here it goes. Get ready to quench your thirst from my fount of wisdom. No man, think for yourself, only get what works for you. It is called foundations, not theory Foundations of computer science (or rather any field of study) are the most important topics you can learn. These lay down the frame of thinking/perspective for that area of study. Yet, I am saddened to hear these called as "theory", and labeled as "unpractical". This couldn't be farther from the truth. Take a look at how I recommend studying distributed systems . Don't you dare call this "theory" and "unpractical". This lays the bedrock that you build your practice on. Don't skimp on the foundations. Don't build your home on quicksand. Keep your hands dirty, your mind cl
Read more >>

Linearizability: A Correctness Condition for Concurrent Objects

-
Image
This paper is from Herlihy and Wing appeared in ACM Transactions on Programming Languages and Systems 1990. This is the canonical reference for the linearizability definition. I had not read this paper in detail before, so I thought it would be good to go to the source to see if there are additional delightful surprises in the original text. Hence, this post. I will dive into a technical analysis of the paper first, and then discuss some of my takes toward the end. I had written an accessible explanation of linearizability earlier; you may want to read that first. I will assume an understanding of linearizability to keep this review at reasonable length. Introduction I love how the old papers just barge in with the model, without bothered by pleasantries such as motivation of the problem. These are the first two sentences of the introduction. "A concurrent system consists of a collection of sequential processes that communicate through shared typed objects . This model encompass
Read more >>

Scalable OLTP in the Cloud: What’s the BIG DEAL?

-
Image
This paper is from Pat Helland, the apostate philosopher of database systems, overall a superb person, and a good friend of mine. The paper appeared this week at CIDR'24. (Check out the program for other interesting papers). The motivating question behind this work is: " What are the asymptotic limits to scale for cloud OLTP (OnLine Transaction Processing) systems? " Pat says that the CIDR 2023 paper "Is Scalable OLTP in the Cloud a Solved Problem?" prompted this question.  The answer to the question? Pat says that the answer lies in the joint responsibility of database and the application. If you know of Pat's work, which I have summarized several in this blog , you would know that Pat has been advocating along these lines before. But this paper provides a very crisp, specific, concrete answer. Read on for my summary of the paper. Disclaimer: This is a wisdom and technical information/detail packed 13-page paper, so I will try my best to summarize the sa
Read more >>

Understanding the Performance Implications of Storage-Disaggregated Databases

-
Image
Storage-compute disaggregation in databases has emerged as a pivotal architecture in cloud environments, as evidenced by Amazon ( Aurora ), Microsoft ( Socrates ), Google (AlloyDB), Alibaba ( PolarDB ), and Huawei (Taurus). This approach decouples compute from storage, allowing for independent and elastic scaling of compute and storage resources. It provides fault-tolerance at the storage level. You can then share the storage for other services, such as adding read-only replicas for the databases. You can even use the storage level for easier sharding of your database. Finally, you can also use this for exporting a changelog asynchronously to feed into peripheral cloud services, such as analytics. Disaggregated architecture was the topic of Sigmod 23 panel . I think this quote summarizes the industry's thinking on the topic. "Disaggregated architecture is here, and is not going anywhere. In a disaggregated architecture, storage is fungible, and computing scales independently.
Read more >>

Designing Data Intensive Applications (DDIA) Book

-
Image
We started reading this book as part of Alex Petrov's book club . We just got started, so you can join us, by joining the discord channel above. We meet Wednesday's 11am Eastern Time.  Previously we had read transaction processing book by Grey and Reuters. This page links to my summaries of that book. Chp 1. Reliable, Scalable, and Maintainable Applications I love the diagrams opening each chapter. Beautiful! The first chapter consists of warm up stuff. It talks about the definitions of reliabilty, scalability, and maintainability. It is still engaging, because  it is written with an educator and technical blogger voice, rather than a dry academic voice. This book came out on 2017. Martin is working on the new version. So if you have comments for things to focus on for the new version, it would be helpful to collect them in a document and email it to Martin. For example, I am curious about how the below paragraph from the Preface will get revised with 8 more years of hindsight:
Read more >>