Sparklyr: how to center a Spark table based on column?
I have a Spark table:
simx
x0: num 1.00 2.00 3.00 ...
x1: num 2.00 3.00 4.00 ...
...
x788: num 2.00 3.00 4.00 ...
and a handle named simX_tbl in the R environment that is connected to this simx table.
I want to do a centering for this table, which is subtracting each column with its column means. For example, calculating x0 - mean(x0), and so on.
So far my best effort is:
meanX <- simX_tbl %>% summarise_all(funs("mean")) %>% collect()
x_centered <- simX_tbl
for(i in 1:789) {
colName <- paste0("x", i-1)
colName2 <- lazyeval::interp(~ a - b, a = as.name(colName), b = as.double(meanX[i]))
x_centered <- x_centered %>% mutate_(.dots = setNames( list(colName2) , colName) )
}
This actually works when I limit the for loop for few iterations (1:5) the x_centered %>% head result is correct. But when I do this for 789 iterations, this error comes out when I try to head it:
Error: C stack usage 7969412 is too close to the limit
Below are the output methods I've already tried that shows the C stack usage error:
x_centered %>% head #show first 6 rows
x_centered %>% select_("x0") #select first column only
x_centered %>% sdf_register("x_centered") #register as table
x_centered %>% spark_dataframe() %>% tbl(sc, "x_centered") #also register as table
spark_write_csv(x_centered, path = "hdfs/path/here") #write as csv
Later on I need to do calculate correlation coefficient for each columns but I don't think I can output with this error.
Is there any way to do this centering correctly / efficiently? I read this question about raising the Cstack limit, but I don't think it's a solution because the data is quite large and there's a risk of overlimit again with bigger data. The actual data is 40GB+ and the data I'm currently using is just a small sample (789 columns x 10000 rows).
Spark version is 1.6.0
EDIT: make title clearer, add tried output methods
Solution 1:
You just use mutate_each / muate_all
library(dplyr)
df <- data.frame(x=c(1, 2, 3), y = c(-4, 5, 6), z = c(42, 42, 42))
sdf <- copy_to(sc, df, overwrite=TRUE)
mutate_all(sdf, funs(. - mean(.)))
Source: query [3 x 3]
Database: spark connection master=local[*] app=sparklyr local=TRUE
x y z
<dbl> <dbl> <dbl>
1 -1 -6.333333 0
2 0 2.666667 0
3 1 3.666667 0
but it looks like it is expanded to a really inefficient (unacceptable for large datasets) window function application. You could be better with more verbose solution:
avgs <- summarize_all(sdf, funs(mean)) %>% as.data.frame()
exprs <- as.list(paste(colnames(sdf),"-", avgs))
sdf %>%
spark_dataframe() %>%
invoke("selectExpr", exprs) %>%
invoke("toDF", as.list(colnames(sdf))) %>%
invoke("registerTempTable", "centered")
tbl(sc, "centered")
Source: query [3 x 3]
Database: spark connection master=local[*] app=sparklyr local=TRUE
x y z
<dbl> <dbl> <dbl>
1 -1 -6.333333 0
2 0 2.666667 0
3 1 3.666667 0
It is not as pretty as dplyr approach but unlike the former one does a sensible thing.
If you want to skip all the invokes you can use dplyr to the same thing:
transmute_(sdf, .dots = setNames(exprs, colnames(sdf)))
Source: query [3 x 3]
Database: spark connection master=local[*] app=sparklyr local=TRUE
x y z
<dbl> <dbl> <dbl>
1 -1 -6.333333 0
2 0 2.666667 0
3 1 3.666667 0
Execution plans:
A helper function (see also dbplyr::remote_query for physical plan):
optimizedPlan <- function(df) {
df %>%
spark_dataframe() %>%
invoke("queryExecution") %>%
invoke("optimizedPlan")
}
dplyr version:
mutate_all(sdf, funs(. - mean(.))) %>% optimizedPlan()
<jobj[190]>
class org.apache.spark.sql.catalyst.plans.logical.Project
Project [x#2877, y#2878, (z#1123 - _we0#2894) AS z#2879]
+- Window [avg(z#1123) windowspecdefinition(ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) AS _we0#2894]
+- Project [x#2877, (y#1122 - _we0#2892) AS y#2878, z#1123]
+- Window [avg(y#1122) windowspecdefinition(ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) AS _we0#2892]
+- Project [(x#1121 - _we0#2890) AS x#2877, z#1123, y#1122]
+- Window [avg(x#1121) windowspecdefinition(ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) AS _we0#2890]
+- Project [y#1122, z#1123, x#1121]
+- InMemoryRelation [x#1121, y#1122, z#1123], true, 10000, StorageLevel(disk, memory, deserialized, 1 replicas), `df`
: +- *Scan csv [x#1121,y#1122,z#1123] Format: CSV, InputPaths: file:/tmp/RtmpiEECCe/spark_serialize_f848ebf3e065c9a204092779c3e8f32ce6afdcb6e79bf6b9868ae9ff198a..., PartitionFilters: [], PushedFilters: [], ReadSchema: struct<x:double,y:double,z:double>
Spark solution:
tbl(sc, "centered") %>% optimizedPlan()
<jobj[204]>
class org.apache.spark.sql.catalyst.plans.logical.Project
Project [(x#1121 - 2.0) AS x#2339, (y#1122 - 2.33333333333333) AS y#2340, (z#1123 - 42.0) AS z#2341]
+- InMemoryRelation [x#1121, y#1122, z#1123], true, 10000, StorageLevel(disk, memory, deserialized, 1 replicas), `df`
: +- *Scan csv [x#1121,y#1122,z#1123] Format: CSV, InputPaths: file:/tmp/RtmpiEECCe/spark_serialize_f848ebf3e065c9a204092779c3e8f32ce6afdcb6e79bf6b9868ae9ff198a..., PartitionFilters: [], PushedFilters: [], ReadSchema: struct<x:double,y:double,z:double>
dplyr optimized:
transmute_(sdf, .dots = setNames(exprs, colnames(sdf))) %>% optimizedPlan()
<jobj[272]>
class org.apache.spark.sql.catalyst.plans.logical.Project
Project [(x#1121 - 2.0) AS x#4792, (y#1122 - 2.33333333333333) AS y#4793, (z#1123 - 42.0) AS z#4794]
+- InMemoryRelation [x#1121, y#1122, z#1123], true, 10000, StorageLevel(disk, memory, deserialized, 1 replicas), `df`
: +- *Scan csv [x#1121,y#1122,z#1123] Format: CSV, InputPaths: file:/tmp/RtmpiEECCe/spark_serialize_f848ebf3e065c9a204092779c3e8f32ce6afdcb6e79bf6b9868ae9ff198a..., PartitionFilters: [], PushedFilters: [], ReadSchema: struct<x:double,y:double,z:double>
Notes:
Spark SQL is not that good in handling wide datasets. With core Spark you usually combine features into a single Vector Column and Spark provides a number of transformers which can be used to operate on Vector data.