Column-Stores vs. Row-Stores: How Different Are They Really?
Proceedings of the 2008 ACM SIGMOD international conference on Management of data, 2008 · paper · Computer science
Compares a commercial row-store, under various configurations, against a column-store on a data warehouse benchmark. Finds the row-store much slower, concluding that both storage layer and query executor must change to obtain column-store benefits.
Checked: PDF from cs.umd.edu, fetched 4 Oct 2026. Page numbers are those of that PDF.
3 findings
Reference or label Reference [11] "G. Graefe. Volcano - an extensible and parallel query evaluation system. 6:120–135, 1994." gives no journal or venue name.
- p. 14 References, [11]
evaluation system. 6:120–135, 1994.
Judge's note: Confirmed on the p.14 image: volume and pages are printed but no venue. The missing venue is visible on the page. Identifying it as IEEE TKDE rests on outside knowledge.
Second check: another model agreed
Report a problem with this finding
Wording "Recall from Section 4 that we experimented with six configurations of System X", but the list that follows has five items and Figure 6 shows five (T, T(B), MV, VP, AI).
- p. 9 Section 6.2, paragraph before numbered list
Recall from Section 4 that we experimented with six configura-
- p. 4 Section 4, first paragraph
We look at three differ- ent classes of physical design
- p. 9 Section 6.2, left column, enumerated list
Recall from Section 4 that we experimented with six
- p. 9 Section 6.2, before numbered list
we experimented with six configura-
Judge's note: On p.9 the numbered list runs 1–5. Figure 6 on p.10 has five bars and columns. Section 4 (p.4) describes three designs plus a 'standard' row-store. The clustered-index VP variant is only a side experiment and is not enumerated. Nothing in the text identifies a sixth configuration.
Second check: another model agreed
Report a problem with this finding
Wording "Materialized views perform best in all cases", but for Q4.2 the traditional design (14.1 s) beats MV (22.4 s) in Figure 6(a), and Figure 5 agrees (RS 14.1 vs RS (MV) 22.4).
- p. 9 Section 6.2, right column
Materialized views perform best in all cases
- p. 10 Figure 6(a), Flight 4 table, Q4.2 row
14.1 25.3 22.4 150.4 280.1
- p. 9 Figure 5, column 4.2
22.4
- p. 10 Figure 6(a), Flight 4 table, row Q4.2
22.4
Judge's note: Read Figure 5 on p.9 and the Figure 6(a) Flight 4 table on p.10. For Q4.2 the values are T 14.1, T(B) 25.3, MV 22.4. Every other query has MV fastest. The universal claim is overstated.
Second check: another model agreed
Report a problem with this finding
19 minor (typos and the like; most checked by the judge only)
Wording The invisible-join gain is summarised three ways: "The invisible join improves performance by 50-75%" (Sec. 6.3.2), "improves performance by about 50%" (Conclusion) and, together with block iteration, "about a factor 1.5 performance gain on average" (Intro). None of these clearly matches the Figure 7(b) averages.
- p. 2 Introduction
offer about a factor 1.5 performance gain on average
- p. 12 Section 6.3.2
The invisible join improves performance by 50-75%
- p. 14 Conclusion
further im-proves performance by about 50%
Only the authors can say what is right here.
Judge's note: Figure 7(b) on p.12 gives tICL 4.0 vs tiCL 7.5 and TICL 6.4 vs TiCL 9.3. Disabling the invisible join raises time by 87.5% and 45% respectively, or cuts time by 47% and 31% if read the other way. The '50-75%' range does not match either reading cleanly, and the conclusion's 'about 50%' differs from it. These are loose summaries, so this is a query for the authors about which baseline the figures use.
Second check: another model agreed
Report a problem with this finding
Wording "Block-processing can improve performance anywhere from a factor of only 5% to 50%": the metric is unclear ('factor of ... %'), and the Figure 7(b) block-processing pairs (tICL 4.0 vs TICL 6.4, tiCL 7.5 vs TiCL 9.3, ticL 14.7 vs TicL 16.0) give 9%–60% as speedups or 8%–37.5% as time reductions. Also "Every query in the SSBM contain" (agreement).
- p. 12 Section 6.3.2, left column
anywhere from a factor of only 5% to 50%
- p. 12 Figure 7(b)
4.0 6.4 7.5 9.3 14.7 16.0 41.0
- p. 2 Introduction, paragraph before contributions
late materialization offers about a factor of 3
- p. 14 Conclusion
late materialization improves performance by a factor of three
Only the authors can say what is right here.
Judge's note: Viewed Figure 7(b) on p.12 and the text. Neither reading of the averages reproduces 5%–50%; per-query values might, but this is not stated. The late-materialization 'factor of 3' (intro p.2, conclusion p.14) against the 41.0/16.0 = 2.56 average is loose rounding. Section 6.3.2 itself says 'almost a factor of three', so I do not treat that part as a separate defect. The grammar slip 'Every query ... contain' is confirmed on p.12.
Second check: another model agreed
Report a problem with this finding
Wording Section 4 says "an optimal set of materialized views for every query flight", but the Section 6.2 list item 5 says "optimal collection of materialized views for every query".
- p. 4 Section 4, Materialized Views
materi- alized views for every query flight in the workload
- p. 9 Section 6.2, list item 5
of materialized views for every query
Judge's note: Checked p.4 (crop C001-b/C012-b) and p.9 (crop C001-c). Section 6.1 also refers to 'the materialized view used for query flight 1', which supports per-flight. Item 5 is a looser wording; either reading is plausible, so this is a minor consistency point.
Second check: another model agreed
Report a problem with this finding
Wording Tuple construction "adds almost a factor of 2", comparing Ticl (Figure 7(b) average 41.0 s) with "CS Row-MV" (Figure 5 AVG 25.9 s); the ratio of the plotted averages is about 1.58.
- p. 13 Section 6.3.2 last paragraph (continued)
it adds almost a factor of 2
- p. 12 Figure 7(b) Ticl
41.0
- p. 9 Figure 5, CS (Row-MV) AVG
25.9
Only the authors can say what is right here.
Judge's note: Viewed Figure 7(b) on p.12 (Ticl 41.0) and Figure 5 on p.9 (CS (Row-MV) AVG 25.9). The text on p.13 explicitly compares these numbers. 1.58 is somewhat short of 'almost 2'. The authors may have used another aggregate; query.
Second check: another model agreed
Report a problem with this finding
Wording Section 9: "All figures containing numbers derived from experiments on the C-Store prototype (Figure 7a, Figure 7b, and Figure 8) have been verified", but Figure 5 also contains C-Store numbers (CS, CS (Row-MV)) and is not listed.
- p. 14 Section 9 Repeatability Assessment
(Figure 7a, Figure 7b, and Figure 8)
- p. 9 Figure 5 rows
CS (Row-MV)
Only the authors can say what is right here.
Judge's note: Viewed p.14 Section 9 and Figure 5 on p.9, which has rows for CS and CS (Row-MV). Figure 5 may have been left out because it mixes in System X results. Even so, 'All figures' does not match the list; query.
Second check: another model agreed
Report a problem with this finding
Wording Section 6.2.1 says VP "performs about as well as the traditional approach (65 seconds versus 43 seconds)", while Figure 6(a) gives Q2.1 T = 43.8 s (rounds to 44) and VP = 65.1 s, about 1.49 times slower.
- p. 10 Section 6.2.1, after SQL
(65 seconds versus 43 seconds)
- p. 10 Figure 6(a), Flight 2, Q2.1
43.8
Judge's note: Viewed Figure 6(a) on p.10: Q2.1 T 43.8, VP 65.1, AI 359.8. The text's '43' truncates 43.8. A difference of about 49% is called 'about as well', although this is plausibly relative to the 360 s of index-only. Wording/rounding point only.
Second check: another model agreed
Report a problem with this finding
Typo In Figure 1 the PART table attribute is misspelled "CATEGOTY"; it should be CATEGORY (the query text uses p.category).
- p. 3 Figure 1, PART table
CATEGOTY
- p. 3 Figure 1
30,0000
Judge's note: The page image (crop C003-a) shows 'CATEGOTY' in the PART box, so this is not an extraction artefact. The query on p.10 uses p.category.
Second check: none (judge only)
Report a problem with this finding
Typo In Figure 1 the CUSTOMER size is printed "Size=scalefactor x 30,0000", a malformed number; presumably 30,000 is meant.
- p. 3 Figure 1, under CUSTOMER table
Size=scalefactor x 30,0000
- p. 3 Figure 1, CUSTOMER size
30,0000
Judge's note: Crop C004-a shows '30,0000' under the CUSTOMER box. The comma grouping is invalid next to the other well-formed sizes (2,000; 6,000,000; 200,000). The intended value is presumably 30,000, consistent with the SSBM spec.
Second check: none (judge only)
Report a problem with this finding
Typo Section 5.4 reads "followed by a join with the data table and the year predicate applied"; this should be "date table".
- p. 6 Section 5.4, first paragraph, left column
followed by a join with the data table and the year predicate applied
- p. 6 Section 5.4, first paragraph
the data table and the year predicate applied
- p. 6 Section 5.4
followed by a join with the data table
Judge's note: Crop C006-a confirms 'data table'. The year predicate is on the date (dwdate) dimension, and there is no 'data' table in the schema.
Second check: none (judge only)
Report a problem with this finding
Typo Subject–verb agreement: "at most one of these two position lists are produced in sorted order" should read "is produced".
- p. 6 Section 5.4
at most one of these two position lists are produced in sorted order
Judge's note: Confirmed in crop C007-a. Grammar only.
Second check: none (judge only)
Report a problem with this finding
Typo A word is missing: "the number values that need to be extracted" should read "the number of values".
- p. 7 Section 5.4.1, left column
the number values that need to be
Judge's note: Confirmed in crop C008-a (p.7, left column).
Second check: none (judge only)
Report a problem with this finding
Typo A verb is missing: "since it not multi-threaded" should read "since it is not multi-threaded".
- p. 8 Section 6.1, right column
since it not multi-threaded
Judge's note: Confirmed in crop C010-a (p.8, right column).
Second check: none (judge only)
Report a problem with this finding
Typo Section 6.1 has "C-Store query executer", while elsewhere the paper writes "executor" (e.g. p.2: "C-Store query executor").
- p. 8 Section 6.1
C-Store query executer
Judge's note: Crop C011-a shows 'executer'. Page 2 (crop C002-a) uses 'executor'. 'Executer' is a variant spelling, so this is an inconsistency only.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 6 caption labels the configuration "AI is all indexes", while the text (Section 4, Section 6.2 item 4, Section 6.2.1) calls it the "index-only" approach.
- p. 10 Figure 6 caption
AI is all indexes
- p. 9 Section 6.2, list item 4
An “index-only” representation
Judge's note: Viewed p.10 (Figure 6 caption 'AI is all indexes') and p.9 (list item 4: 'An "index-only" representation'). This is the same configuration under two names. The legend defines AI, so readers can match them, but the naming is inconsistent. Cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo Section 6.3: "the other systems does not need join together columns from the same table" should read "do not need to join together".
- p. 11 Section 6.3, right column top
the other systems does not need join together columns
- p. 11 Section 6.3, right column top
the other systems does not need join together
- p. 11 Section 6.3
does not need join together columns
Judge's note: Confirmed on the p.11 image, right column top: the subject-verb agreement is wrong and 'to' is missing.
Second check: none (judge only)
Report a problem with this finding
Typo Section 6.3.2: "returns a pointer to an array than can be iterated through directly" should read "that can".
- p. 11 Section 6.3.2, last paragraph
returns a pointer to an array than can be iterated through di-
- p. 11 Section 6.3.2, last paragraph
an array than can be iterated
Judge's note: Confirmed on the p.11 image, last paragraph of the right column.
Second check: none (judge only)
Report a problem with this finding
Typo Section 6.3.2: "since such are tuples immediately discarded" should read "since such tuples are immediately discarded".
- p. 12 Section 6.3.2, right column
since such are tuples immediately discarded
Judge's note: Confirmed on the p.12 image, end of the late-materialization paragraph in the right column.
Second check: none (judge only)
Report a problem with this finding
Typo Section 6.3.3: "only those columns relevant for a query need to read in" should read "need to be read in".
- p. 13 Section 6.3.3, left column
relevant for a query need to read in
Judge's note: Confirmed on the p.13 image, left column, Section 6.3.3 first paragraph.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [24]: "Buffering databse operations for enhanced instruction cache performance" should read "database".
- p. 14 References, [24]
Buffering databse operations for
- p. 14 References, [24]
Buffering databse operations
Judge's note: Confirmed on the p.14 image, reference list.
Second check: none (judge only)