Storage Frontend Postmortem; Google incident #123,  JaneCorp incident #22  Date :  2017­10­26  Authors :  Adrian H, Gwendolyn S, Dave R (Google);  Jane D (JaneCorp)  Status :  Complete, action items in progress  Summary :  New storage frontend release caused increased end­user latency for uncached  objects of a common size in a single region.  Background:  Google’s storage frontend service (“storage FE”) handles all customer read/write requests for  binary objects for the Google Cloud Blob storage product. It is deployed on a regional basis,  with one pool of frontends in each Cloud region and no cross­region dependencies. It is  managed by the Storage SRE team who control configuration pushes and binary releases to the  service. The frontends in each region share a cache to reduce the traffic load on the underlying  storage system. Over 70% of read requests are cache hits. Object sizes read through storage  FE vary from 1KB to 1MB.     The release process for storage FE happens weekly. There is a deployment to a single Cloud  region (us­arctic1) on Day 1, before it goes to 2 other regions on Day 2, and all regions by Day  4. This is to mitigate the impact of a bad release. Before release there is an integration testing  process that aims to identify significant performance regressions.    Monitoring and alerting of storage FE is broken down by Cloud region, so there is a distinct  monitoring console and alerting configuration for us­arctic1. Monitoring includes the cache hit  rate and overall read latency, but there is not a breakdown of latency by object size.    Cloud customer JaneCorp uses Cloud Blob as a backend for their JaneApp application. This  has a 99.9% availability target and 1000ms @ 95% latency SLO. Most of their requests are  cache misses in storage FE, and they use objects in the 4KB­32KB size range. They serve from  two regions: us­arctic1 and europe­iceland1.    Impact :   ● 3% of all storage FE read requests (20% of cache misses) in us­arctic1 region served  with out­of­SLO latency for 2.5 hours.  ● Customer app JaneApp served 50% of requests with out­of­SLO latency for 2 hours,  burning 1.5x their 30 day error budget.    Root Causes :  ● incorrect implementation of a new object lookup function;   ● bug only manifested on uncached objects of a particular size;  

● ●

qualification tests did not separately measure latency impact by object size and cache  status;  production monitoring did not measure or alert on increased latency by object size or  cache status. 

  Trigger : Binary rollout.    Resolution : Rolled back binary and verified that latency returned to normal. Identified and  reverted offending code change.    Detection : Customer monitoring.    Action items :    Action Item 

Type 

Owner 

Bug / status 

Extend storage FE integration tests to  break down latency by cache status  and object size bucket 

prevent 

storage dev  (Google) 

b­12345 /  complete 

Implement per­region storage FE  latency monitoring broken down by  cache status and object size bucket. 

detect 

storage SRE  (Google) 

b­23456 /   in progress 

Update JaneApp Ops playbook with  accurate instructions for a regional  failover. 

mitigate 

Ops  (JaneCorp) 

MC­54321 /  complete 

Implement and document thru­stack  latency breakdown for JaneApp 

detect 

JaneApp dev  (JaneCorp) 

MC­54322 /  in progress 

Schedule monthly region failover drills  for JaneApp. 

mitigate 

Ops  (JaneCorp) 

MC­54323 /  not started 

  Lessons learned    What went well:  ● Storage FE rollback was quick once the problem was identified.  ● The shared monitoring for JaneApp enabled Google Cloud Support to quickly confirm  the latency problem and location and escalate.  ● The 1­region­per­day rollout schedule limited the impact of the bug to a single region.  ● JaneApp monitoring and alerting paged the on­call for high service latency within 10  minutes of the rollout.  ● JaneCorp Ops had an existing plan to fail out of a region as mitigation for a single­region  Google service problem.    What went badly: 

● ● ● ● ● ● ●

Storage FE Integration / qualification testing did not detect the problem. It does measure  latency of the new binary, but does not break this down by object cache status or size.  Storage FE per­region production monitoring did not clearly show the problem, similarly  due to too­broad aggregation.  No alert fired for the affected region.  Customer detected the problem (rather than Google).  JaneApp monitoring did not show thru­stack latency breakdown, making it hard to  identify the source of the latency.  JaneApp was out of latency SLO for 50% of their traffic for 2 hours, blowing their 30 day  error budget.  JaneApp took 30 minutes to fail out of the affected region due to stale instructions in the  Ops playbook.  

  Where we got lucky:  ● Only reads of objects between 4KB and 32KB in size were materially affected, and only  uncached objects.  ● JaneApp was sufficiently affected in the one region to detect the problem and alert  Google before it went to other regions.  ● The incident happened during JaneCorp business hours, when developers were on  hand to help debug.    Timeline  All times Pacific.  2017­10­19  10:14 Storage FE developer commits change c­12345 improving object lookup reliability.   This contains a bug that makes reads of objects of size 4KB­32KB take 3x as  long as normal to retrieve.  17:00 Change c­12345 incorporated into release branch r­777  2017­10­23  13:15 Integration testing of r­777 completes. Overall read latency is increased by about   1% but this is not perceptible.  2017­10­26  08:00 Storage FE oncall triggers the rollout of release 555 of Storage FE to the Day 1  region (us­arctic1)  08:10 us­arctic1 storage FE pool starts serving with the new release    08:20 us­arctic1 overall storage FE read latency increases by 2% from baseline [graph]  08:25 JaneApp pages oncall for out­of­SLO latency in us­arctic1 region    08:40 JaneApp oncall starts diagnosis work  09:10 JaneApp oncall can’t identify latency source from available monitoring, contacts  JaneApp developer  09:40 JaneApp developer and oncall confirm that the latency is coming from Cloud Blob  service.  09:42 JaneApp oncall announces intention to fail JaneApp out of us­arctic1, starts 

following the playbook instructions  09:50 Playbook instructions aren’t working, traffic is still flowing to us­arctic1  10:05  JaneApp secondary oncall  raises P1 Google Support ticket #67890: “High latency  for Cloud Blob in us­arctic1 region”  10:10 JaneApp successfully failed out of us­arctic1    10:18 Google Cloud Support engineer confirms on ticket that they see the problem in the  JaneApp shared monitoring, pages storage FE oncall  10:25 storage FE oncall acks page, starts investigation  10:35 storage FE oncall can’t see any clear latency regression in us­arctic1, starts  digging into logs.  10:45 storage FE oncall identifies that uncached objects might have some latency  regression. [graph]   10:50 storage FE oncall has sufficient suspicion of the recent rollout and starts a rollback  of release 555 in us­arctic1.  11:10 us­arctic1 storage FE pool now serving with original release    13:30 Change c­12345 identified as suspicious.  14:15 Integration testing re­run with object size breakdown, confirms regression.  14:30 Change c­12345 rolled back.      

Storage Frontend Postmortem Services

Oct 19, 2017 - Google's storage frontend service (“storage FE”) handles all customer read/write requests for binary objects for the ... Cloud customer JaneCorp uses Cloud Blob as a backend for their JaneApp application. This ... Customer app JaneApp served 50% of requests with outofSLO latency for 2 hours, burning ...

113KB Sizes 1 Downloads 327 Views

Recommend Documents

Pfff, a PHP frontend
Page 1. Pfff, a PHP frontend. Yoann Padioleau [email protected]. February 14, 2010. Page 2. Copyright cс 2009-2010 Facebook. Permission is ...

Title: RF Frontend -
If RFFE_CLK supplied, omit C26 C27 X4. C26 C27 X4 only if layout allows as option for inexpensive local clock with worse phase noise. Clock Build Options.

Compilers: Backend to Frontend - GitHub
Sep 17, 2006 - Line 3 says that the label scheme_entry should be considered a ...... Figure 1.10: One way of implementing proper tail calls is by collapsing.

Frontend-Intern-Atameo.pdf
Page 1 of 1. Frontend-Intern-Atameo.pdf. Frontend-Intern-Atameo.pdf. Open. Extract. Open with. Sign In. Main menu. Displaying Frontend-Intern-Atameo.pdf.

TRAINABLE FRONTEND FOR ROBUST AND ... - Research
Google, Mountain View, USA. {yxwang,getreuer,thadh,dicklyon,rif}@google.com .... smoother M, PCEN can be easily expressed as matrix operations,.

Uncle Bob's Self Storage used AdWords Location ... Services
Uncle Bob's Self Storage, owned and operated by Sovran Self Storage, Inc., ... use their service. They were experiencing a very high advertising Cost per. Acquisition and were looking for a better solution. “ When we started doing location targetin

pdf-0946\forensic-taphonomy-the-postmortem-fate-of-human ...
... one of the apps below to open or edit this item. pdf-0946\forensic-taphonomy-the-postmortem-fate-of-human-remains-forensicnetbase-from-crc-press.pdf.

react-frontend-at-strv-jirikomon.pdf
Connect more apps... Try one of the apps below to open or edit this item. react-frontend-at-strv-jirikomon.pdf. react-frontend-at-strv-jirikomon.pdf. Open. Extract.

Altered expression of mitochondria-related genes in postmortem ...
Nov 24, 2004 - ... following evidence: (i) altered energy metabolism in brains of patients with BD, as ..... pH level are colored in dark green. Asterisks indicate ...

Altered expression of mitochondria-related genes in postmortem ...
Nov 24, 2004 - We performed large-scale DNA microarray analysis of postmortem brains of patients with ... There are several issues of concern about interpreting data from clinical ... Among the 676 mitochondria-related probe sets (mt-probe.

Storage router and method for providing virtual local storage
Jul 24, 2008 - Technical Report-Small Computer System Interface-3 Generic. PacketiZed Protocol ... 1, 1996, IBM International Technical Support Organization, ..... be a rack mount or free standing device With an internal poWer supply.

Storage-based Intrusion Detection: Watching storage ...
Section 5 describes a prototype storage IDS embedded in an NFS server. Sec- ..... For small numbers of dedicated servers in a machine room, either approach is ...

TRAINABLE FRONTEND FOR ROBUST AND ... - Research at Google
tion, we introduce a novel frontend called per-channel energy nor- malization (PCEN). ... In this section, we introduce the PCEN frontend as an alternative to the log-mel .... ing background noises and reverberations, where the noise sources.

Unified Storage
Dec 10, 2009 - from any web browser, PC or mobile device. Online and Offline – synchronization ensures information access even without Internet access. Search and Organize – find, tag and create information. Share and Publish – securely share i

Storage of materials
Mar 24, 1992 - (74) Attorney, Agent, or F irmiNeifeld IP Law, PC. (57). ABSTRACT .... 1989”; pp. 234229. Fax from Eric Potter Clarkson to EPO dated Oct. 15, 1999 ..... substances Which in their free states are extremely labile are found to ...

Storage of materials
Aug 28, 2001 - 4,847,090 A * 7/ 1989 Della Posta et al. ...... .. 424/440 ... Water-soluble or sWellable glassy or rubbery composition .... drate Water Systems”.*.

Universal storage management system
Jul 31, 2002 - 1995) pp. 1- 126. Leach, P., et al., “CIFS: A Common Internet File System,” Microsoft ..... glossary and related article on open support) (Special Report: Fault ..... “Encore Sets Agreement with Bell Atlantic Business Systems Ser

Storage router and method for providing virtual local storage
Jul 24, 2008 - CRD-5500, Raid Disk Array Controller Product Insert, pp. 1-5. 6'243'827 ..... Data Book- AIC-1 160 Fibre Channel Host Adapter ASIC (Davies Ex. 2 (CNS ..... devices 20 comprise hard disk drives, although there are numerous ...

Universal storage management system
Jul 31, 2002 - 5,568,628 A 10/1996 $61611 6161. 4,467,421 A ... 10/1997 Yamarnoto et al. .... Veritas. Software Corp. and Veritas Software Global Corporation, Civil ...... Companies, Approaches to RAID and IBM Storage Product Plans;.

3 STORAGE BATIERY
Sep 5, 2008 - storage components typically use a battery-centric topology. This is because the battery provides a stable voltage and ... from a utility grid at unity power factor, and a renewable energy port that sources energy into the .... of the A

Storage of materials
Aug 28, 2001 - Declaration under 37 CFR 1.132 re: Franks et al. Ser. No. 08/241,457 ..... continue it for some hours, for instance 24 to 36 hours. As evaporation ...