Dataset · public postmortems · SoCC 2016

355 of 597 public cloud outages have an unknown root cause

Gunawi et al., SoCC 2016, Table 3, mark 355 of 597 unplanned outages as UNKNOWN. The study covers 32 popular Internet services and 1,247 public reports from January 2009 through December 2015. The paper states that only 40% of those descriptions reveal a root cause and only 24% reveal a fix. Among known root-cause tags, the three largest counts are UPGRADE 54 (16%), NETWORK 52 (15%), and BUGS 51 (15%). STORAGE is 13 occurrences on 4 services (4% of known causes). HARDWARE is 5 occurrences on 4 services (1% of known causes). One outage can carry more than one tag, so the occurrence column is not a partition of the 597 outages. In the yearly column, M means at least 10, not exactly 10.

File /data/datasets/postmortem-unknown-causes.csv · JSON metadata · All datasets · what fails in storage subsystems

Method

The CSV transcribes stated figures from Gunawi et al., SoCC 2016. It does not digitize a chart and it does not refit a percent. The population is the paper's cloud outage study: 32 Internet services named in Table 1, searched as "serviceName outage month year" on Google and Bing, first 30 hits, every month from January 2009 through December 2015. That search produced 1,247 unique links describing 597 unplanned outages. Each outage was read by at least four authors. A vague "configuration problem" is tagged CONFIG only. The authors say they do not speculate.

Table 3 is the table3 block. Occurrences (count) are the Cnt column. Services with this root cause (count) are the #Sv column. Share (%) is the paper's percent among known root causes. The caption says the percent is not computed on the UNKNOWN row, so that share cell is empty. One outage can carry more than one root-cause tag, so the occurrence column is not a split of the 597 outages. Yearly marks are the seven-token string for 2009 through 2015, copied as printed. M means at least 10. A dash stays a dash.

The uptime rows are section 3.2 prose, not a reading of Figure 1 or Figure 2. The paper's 99% line is not more than 88 hours of annual downtime. The 99.9% line is not more than 8.8 hours. Those ceilings are definitions, not measured downtimes, and they stay in the signal text rather than in a count column. Impact rows and fix rows are the percents printed in section 6. They are not Figure 5 cell counts. These figures are the paper's counts of public reports, not a provider's advertised uptime.

Limits

• Population and date: 32 popular Internet services, public headline news and post-mortem reports only, January 2009 through December 2015. Not an industry sample. Not a vendor fleet. Unreported outages are outside the study, and the paper calls the availability counts minimums.
• Not measured here: the text of the 1,247 links, per-service uptime under real names, and any outage before 2009 or after 2015. Planned maintenance and failures of non-essential operations are excluded by the paper's definition.
• Not measured here: exact yearly counts wherever Table 3 prints M. Seven tens would be 70, not the 355 UNKNOWN occurrences. This file does not solve an M even when the other digits leave a single residual.
• Not measured here: Figures 1, 2, 3, 4, and 5. Whisker medians, bar heights, and the Figure 5 grids are not in the CSV. Section 6's prose percents are included. The impact percents and the fix percents each sum to 99 as printed.
• Not measured here: a line rate, a throughput, a disk AFR, or a byte count. STORAGE is a tag on a public report, 13 occurrences, not a drive failure rate. The Jiang et al. FAST 2008 storage-subsystem counts are a different study, linked below, and are not added to these 13.
• The PDF says COSDB, 597 descriptions, 1,247 links, and 3,249 tags, is downloadable from reference [1], http://ucare.cs.uchicago.edu/projects/cbs/. A fetch of that URL on 2026-10-09 did not return a verified page. This file does not describe what that URL hosts, and it does not redistribute the reports.
• Murat Demirbas, 6 November 2016, repeats the 597, 1,247, 355, 40%, and 24% sentences. That review writes the search through December 2006. The paper's methodology sentence is January 2009 through December 2015. Rows in this file cite the paper, not the review. The average-year uptime sentence says "across the six years" inside a seven-year window. That phrase is kept as printed.

Counted here

From the Table 3 cells, not from a total the paper prints: the 13 known-cause occurrence counts sum to 340. Adding the 355 UNKNOWN occurrences gives 695 tag occurrences. That is not 597. The paper says an outage can be tagged with more than one root cause. 597 minus 355 is 242. That complement is not a printed cell. The paper states the disclosure rate as 40%, and this file does not replace 40 with a ratio of those two counts.

Where a yearly string contains no M, reading each dash as zero makes the seven numbers sum to the occurrence count. That check covers LOAD, POWER, SECURITY, HUMAN, STORAGE, SERVER, NATDIS, and HARDWARE. It is a check, not a second measurement, and the CSV still stores the dash. Where M appears, the hidden year is not filled in.

Comparison

HARDWARE and SERVER both use section 5.11 because both sit under "Miscellaneous Server and Hardware Failures." NATDIS is section 5.12. The Jiang et al. comparison of disk and interconnect events is a separate dataset: what-fails-comparison.

Columns

row_kind (string)

Which stated cut the row transcribes: population, uptime, table3, impact, or fix.

measure (string)

The paper's name for the row: a study total, an uptime cut, a Table 3 root-cause tag, an impact tag, or a fix tag.

section (string)

Section number as printed. Empty only for UNKNOWN, which has a blank section cell in Table 3. HARDWARE and SERVER are both 5.11.

outages_count (integer)

Unplanned outages (count). Filled only on the population row for the 597 outages. Empty means the paper does not give this row as an outage total.

reports_count (integer)

Public reports (count). Filled only on the population row for the 1247 unique links. Empty otherwise.

tags_count (integer)

Outage metadata tags (count). Filled only on the population row for the 3249 COSDB tags. Empty otherwise. Not a sum of Table 3.

study_services_count (integer)

Services in the study (count). Filled only on the population row for the 32 services in Table 1. Empty otherwise.

tagged_services_count (integer)

Services with this root cause (count). Table 3 column #Sv. Empty when the row is not a Table 3 root cause.

services_short_of_uptime_count (integer)

Services short of the stated uptime (count), from section 3.2. Empty when the row is not an uptime cut.

occurrences_count (integer)

Root-cause occurrences (count). Table 3 column Cnt. Not a partition of the 597 outages. Empty outside Table 3.

share_percent (integer)

Printed share (%). The basis is share_basis, not a single denominator. Empty when the paper prints a dash or does not print a percent. Table 3 UNKNOWN is empty.

share_basis (string)

What share_percent is a percent of, using the paper's wording. Empty when share_percent is empty.

yearly_marks (string)

Table 3 yearly marks for 2009 through 2015, seven tokens separated by periods, copied as printed. M means at least 10, not exactly 10. A dash is the paper's dash, not a zero stored by this file. Empty outside Table 3.

signal (string)

The report wording or column the paper uses to assign this tag or count.

misses (string)

What that signal does not prove, from a sentence in the same paper or from a blank the paper leaves.

citation_id (string)

Id into the citations list. Open that URL for the sentence. Every row cites the SoCC 2016 PDF.

License

Small derived table of figures stated in Gunawi et al., SoCC '16, with credit. Not a Creative Commons license. The first page of the PDF that was opened says abstracting with credit is permitted. It also says that to copy otherwise, or to republish, post on servers, or redistribute to lists, requires prior specific permission and/or a fee, requested from permissions@acm.org. Copyright is held by the owner/author(s). Publication rights are licensed to ACM. DOI as printed: 10.1145/2987550.2987583. This file is not the paper, not the 1247 reports, and not COSDB. The paper says COSDB is downloadable from reference [1], http://ucare.cs.uchicago.edu/projects/cbs/. A fetch of that URL on 2026-10-09 did not return a verified page.

Canonical SoCC 2016 PDF used for this table · DOI as printed on that PDF: 10.1145/2987550.2987583

Sources

1. Why Does the Cloud Stop Computing? Lessons from Hundreds of Service Outages (accessed 2026-10-09)
2. Paper Review: Why Does the Cloud Stop Computing? Lessons from Hundreds of Service Outages (accessed 2026-10-09)