Dataset · SSD failures · SYSTOR 2016

A 2016 SMART predictor reached 98% recall on one Seagate model but 81% on a Hitachi model, scored with random splits on Backblaze data

Botezatu et al., KDD 2016, predicted disk replacement from SMART data of 50,984 Backblaze disks: 98% recall on one Seagate model and 81% on one Hitachi model, scored with 100 random 80/20 splits. Cross-check rows come from Azure (ATC 2018), Backblaze (2016), and a data center operator (FAST 2020).

Download CSV15 rows · 7 columns · CSV

File /data/datasets/smart-replacement-prediction-random-splits.csv · JSON metadata · All datasets · Human-readable note on hesela.com

Method

The CSV transcribes figures printed in Narayanan et al. (SYSTOR 2016) and two cross-check rows from NetApp (FAST 2020) and Backblaze (2016). It does not read chart values. AFR is the share of devices with failures divided by device years. A failure here is a fail-stop that takes a server down.

Limits

Two disk families from one public fleet, 2013 to 2015; a model is needed per manufacturer.
Scores come from random 80/20 splits of a downsampled set, not from a split by time.
The paper does not say the test set keeps the true 2.5% to 3% share of replaced disks.
The Backblaze post is the same fleet's own analysis, not an independent dataset.
The cross-check rows use other data and other definitions.

Table

15 rows, the same rows as the CSV. Value low and value high are the printed range ends, or a single printed value in the high column. Units are in the unit column. Sources are the pages opened on 2026-10-11.
Measure	Value low	Value high	Unit	Scope	Note	Citation
Disks in the Backblaze dataset used	not stated	50984	disks	Backblaze public daily SMART logs	Only a Hitachi and a Seagate family were kept; other makers had too few samples or too few SMART values.	botezatu-kdd16-pdf
Months of data kept	not stated	17	months	April 2013 to June 2015 collection; first months dropped	Over 70% of SMART values were not collected before January 2014.	botezatu-kdd16-pdf
Replaced share of the two main models	2.5	3	percent of disks	Seagate ST4000DM000 (SgtA) and Hitachi HDS722020ALA330 (HitA)	The healthy class was downsampled to 1,000 (SgtA) and 500 (HitA) before training.	botezatu-kdd16-pdf
Error on the best Seagate model	1	2	percent over 100 runs	SgtA, regularized greedy forest	Printed as 98% accuracy and 1-2% error.	botezatu-kdd16-pdf
Recall for replaced disks, Seagate model	not stated	98	percent	SgtA, 100 random 80/20 splits	Median over runs.	botezatu-kdd16-pdf
Recall for replaced disks, Hitachi model	not stated	81	percent	HitA, 100 random 80/20 splits	Precision, recall, and F-score for Hitachi are 14 to 19% lower than for Seagate.	botezatu-kdd16-pdf
Recall of a simple decision tree on a small set of SMART attributes	not stated	53	percent	SgtA; 44% for HitA	The authors' baseline with a commonly used subset of attributes.	botezatu-kdd16-pdf
Replaced Seagate disks predicted 10 days ahead	not stated	92	percent	SgtA; 97% at 3 days	From a snapshot taken 10 days before replacement.	botezatu-kdd16-pdf
Replaced Seagate disks predicted 30 days ahead	not stated	73	percent	SgtA; HitA 75% at 30 days	From a snapshot taken 30 days before replacement.	botezatu-kdd16-pdf
Evaluation repeats with random training and test splits	not stated	100	splits of 80% training and 20% test	Same models	Not split by time.	botezatu-kdd16-pdf
True positive rate with random cross-validation at a 0.1% false positive rate	not stated	91.64	percent	Microsoft Azure, Dataset 1	Xu et al. cite Botezatu et al. among the studies that use cross-validation.	xu-atc18-pdf
True positive rate with online prediction at the same false positive rate	not stated	36.5	percent	Same dataset and model, training before testing in time	The drop is the authors' evidence that random splits are optimistic.	xu-atc18-pdf
Failed drives with at least one of five SMART counts above zero	not stated	76.7	percent of failed drives	Backblaze, 67,814 drives, 2016	23.3% showed no warning; the post is the fleet's own analysis, not an independent dataset.	backblaze-smart-2016
Operational drives with at least one of the five SMART counts above zero	not stated	4.2	percent of operational drives	Same fleet	Counts above zero alone are a weak alarm.	backblaze-smart-2016
Matthews correlation coefficient with 5-fold cross-validation	not stated	0.95	correlation coefficient	disks of a leading data center operator, 10-day lead time	Also a random partition, so the same caution applies.	lu-fast20-pdf

Columns

measure (string): The quantity as the source names it.
value_low (number): Lower end of a printed range. Empty when the source prints a single value.
value_high (number): Single printed value, or the upper end of a range.
unit (string): Unit of the value columns: disks, months, percent, splits, or correlation coefficient.
scope (string): Population and window the value applies to.
note (string): What the value is not, or the source wording behind it.
citation_id (string): Id of the opened source in the citations list.

License

Small derived table of figures printed in Botezatu et al. (KDD 2016), Xu et al. (ATC 2018), a Backblaze post of 6 October 2016, and Lu et al. (FAST 2020), with credit. Not a Creative Commons license. The papers remain with their authors or publishers. This file is not a copy of the papers and not the data.

Sources

Predicting Disk Replacement towards Reliable Data Centers, Botezatu, Giurgiu, Bogojeska, and Wiesmann, ACM SIGKDD 2016, pages 39 to 48 (KDD PDF) (accessed 2026-10-11)
Improving Service Availability of Cloud Systems by Predicting Disk Error, Xu et al., USENIX ATC 2018 (USENIX PDF, section 3 on online prediction) (accessed 2026-10-11)
What SMART Stats Tell Us About Hard Drives, Backblaze, 6 October 2016 (accessed 2026-10-11)
Making Disk Failure Predictions SMARTer!, Lu et al., USENIX FAST 2020 (USENIX PDF, section 4) (accessed 2026-10-11)