Amazon Aurora: Design Considerations for High Throughput Cloud-Native Relational Databases
Proceedings of the 2017 ACM International Conference on Management of Data, 2017 · paper · Computer science
Describes the architecture of Amazon Aurora, a cloud relational database that pushes redo processing into a scale-out storage service to reduce network traffic. This enables fast recovery and data-loss-free failover; it also shares lessons from over 18 months in production.
Checked: PDF from web.stanford.edu, fetched 4 Oct 2026. Page numbers are those of that PDF.
4 findings
Substance Section 4.2.3 states a page is evicted only "if its “page LSN” ... is greater than or equal to the VDL", yet the stated guarantee (a) that "all changes in the page have been hardened in the log" requires page LSN less than or equal to the VDL.
- p. 6 Section 4.2.3 Reads, 2nd paragraph
if its “page LSN” (identifying the log record associated with the latest change to the page) is greater than or equal to the VDL
- p. 6 Section 4.2.3, second paragraph
is greater than or equal to the VDL
Judge's note: Read on the page 6 image. VDL is the highest durable point, and records above it can be truncated. A page whose latest change has LSN > VDL therefore holds non-durable changes. Allowing eviction when page LSN >= VDL would permit exactly those pages to be evicted, contradicting guarantee (a) in the next sentence. It would also forbid evicting fully durable pages with LSN < VDL. The replica rule in 4.2.4 uses "LSN is less than or equal to the VDL" for the same durability notion. The inequality as printed is reversed.
Second check: another model agreed
Report a problem with this finding
Substance Section 6.1.1: MySQL 5.7 "tops out at 20,000 reads/sec and 125,000 writes/sec"; the read and write figures are swapped (should be about 125,000 reads/sec and 20,000 writes/sec).
- p. 8 Section 6.1.1, paragraph below Figure 6
which tops out at 20,000 reads/sec and 125,000 writes/sec
- p. 8 Section 6.1.1, last paragraph of left column
tops out at 20,000 reads/sec and 125,000 writes/sec
Judge's note: In Figure 7 (write-only), the MySQL 5.7 bar at r3.8xlarge is about 15K writes/sec. In Figure 6 (read-only), the MySQL 5.7 bars at the largest sizes are about 120-140K reads/sec. As printed, MySQL would exceed Aurora's 121,000 writes/sec, which contradicts the '5x' claim in the same sentence. Swapped, the ratios are about 4.8x and 6x.
Second check: another model agreed
Report a problem with this finding
Reference or label Section 8 cites "a recovery protocol based on ARIES [5]", but [5] is Mohan & Lindsay, "Efficient commit protocols for the tree of processes model of distributed transactions"; ARIES is reference [7].
- p. 11 Section 8, Recovery paragraph
based on ARIES [5]
- p. 11 Reference [5]
Efficient commit protocols for the tree of processes model of distributed transactions
- p. 6 Section 4.3
ARIES [7]
- p. 6 Section 4.3
such as ARIES [7]
Judge's note: In the page 11 reference list, [7] is 'ARIES: A transaction recovery method...' and [5] is the commit-protocols paper. Section 4.3 on page 6 cites 'ARIES [7]' correctly.
Second check: another model agreed
Report a problem with this finding
Reference or label Section 6.2.2: "The median (P50) and 95th percentile (P99) latencies" labels the 95th percentile as P99; the following text and the Figure 9/10 captions use P95.
- p. 9 Section 6.2.2, first paragraph
95th percentile (P99)
- p. 9 Section 6.2.2 next paragraph / Figure 9 caption
the P95 latencies ranged between 40ms to 80ms
- p. 9
The median (P50) and 95th percentile (P99) latencies
Judge's note: The page 9 image shows '(P99)'. The next paragraph says 'the P95 latencies', and both captions read 'P50 vs P95'.
Second check: another model agreed
Report a problem with this finding
10 minor (typos and the like; most checked by the judge only)
Wording "Aurora’s performance doubles for each higher instance size" is not borne out at the last step of Figures 6 and 7 (r3.4xlarge to r3.8xlarge is about 1.6x).
- p. 8 Section 6.1.1, text below Figure 6
Aurora’s performance doubles for each higher instance size
Judge's note: Bar heights read from the page 8 image: Fig. 7 Aurora about 75K at 4xlarge and 121K at 8xlarge; Fig. 6 about 370K and 600K. The earlier steps are roughly 2x. The last step is about 1.6x, which is well beyond reading error. This is an overstatement in the wording.
Second check: another model agreed
Report a problem with this finding
Typo Misspelling in Section 2.1: "concurrently occuring background noise failures" should read "occurring".
- p. 2 Section 2.1
concurrently occuring background noise failures
Judge's note: The crop of page 2 (right column) shows "occuring" as typeset.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.2: "multiple replicas instances" should read "multiple replica instances".
- p. 4 Section 3.2
multiple replicas instances
Judge's note: The crop of page 4 shows "multiple replicas instances deployed across multiple AZs". Later in the same paragraph the paper writes "the replica instances".
Second check: none (judge only)
Report a problem with this finding
Typo On page 8, Figure 7 (write-only) is placed above Figure 6 (read-only), and the text cites "Figure 7 and Figure 6"; the figures are numbered out of order of appearance.
- p. 8 Figure captions, left column
Figure 7: Aurora scales linearly for write-only workload
- p. 8 Section 6.1.1
The results are shown in Figure 7 and Figure 6
Judge's note: The page 8 image confirms the layout order. The pointers resolve correctly, so this is cosmetic only.
Second check: none (judge only)
Report a problem with this finding
Typo Section 6.2.3: "can “can cause strange bugs”" duplicates the word 'can'.
- p. 9 Section 6.2.3
can “can cause strange bugs”
Judge's note: The page 9 image shows both occurrences of 'can'.
Second check: none (judge only)
Report a problem with this finding
Typo Section 6.2.1: "the average response time 5.5 ms, a 3x improvement" lacks a verb ('was').
- p. 9 Section 6.2.1
the average response time 5.5 ms, a 3x improvement
- p. 9 Section 6.2.1
the average response time 5.5 ms
- p. 9 6.2.1
a 3x improvement
Judge's note: The page 9 image confirms the missing verb. The '3x' is a rounding of 15/5.5 = 2.73, consistent with Figure 8's 'Aurora 3X faster' label; I do not count that as a separate defect.
Second check: none (judge only)
Report a problem with this finding
Typo Section 7.1: "when it is unlikely that all of their customers active at once" lacks the verb 'are'.
- p. 9 Section 7.1
all of their customers active at once
Judge's note: The page 9 image confirms the wording.
Second check: none (judge only)
Report a problem with this finding
Typo Section 7.1: "report ha ving more than 50,000" splits 'having' with a stray space.
- p. 9 Section 7.1, last sentence
report ha ving more than 50,000
Judge's note: The page 9 image shows the gap in the typeset text, so it is not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo References [4] ("...database management system”.") and [26] ("...data processing tool”.") end their titles with a closing quotation mark that has no opening mark.
- p. 11 Reference [4]
management system”.
- p. 12 Reference [26]
data processing tool”.
- p. 11 Reference [4]
database management system”.
Judge's note: The page 11 image (ref [4]) and the page 12 crop (ref [26]) both show the unmatched closing quotes.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [26] reads "MapReduce: a flexible data processing tool”.", with a closing quotation mark that has no matching opening mark.
- p. 12 Reference [26]
data processing tool”.
Judge's note: The crop of page 12 shows a closing curly quote after "tool" in reference [26]. No opening quote appears before "MapReduce". This is a cosmetic punctuation slip.
Second check: none (judge only)