Keep R Model in memory for Rest API - r

Question

We have a GLM R model that is ~ 2 GB in size. We are using this model to service a REST API. We need < 3 second response times. The problem is that using the conventional saveRDS / readRDS functions takes way too long. Our last option is to find a way to keep the R Model in memory in between sessions.
At a high level:
Service a REST API.
Keep a multi gigabyte model in memory.
Keep response times below 3 seconds.
Stuff we've tried and doesn't work:
Shrinking the size of the model. Our data scientist says that's as small as he can get it.
I've experimented with compression and other settings on saveRDS. The best I could do is 12 seconds to load the model from file.
We tried Microsoft R Server. All the web API requests would be redirected to the same session. Problem is that we had to wrap a decent amount of code around keeping the session alive. Even then it would flake out often.
Microsoft R Server Real Time is out because it only accepts models generated by the ScaleR. I know ScaleR has a GLM function, but I've been told its not an option.
Faster IO doesn't seem to help. It appears that the bottleneck is the deserialization of the rData file. R being single threaded doesn't help.
Edit:
The question is what R for REST API library / service would allow us to statefully keep a model in memory in between calls.

Answer 1

Continuing on my comment from above as well as the suggestion of #TenniStats, the best approach is to reduce the size of the GLM. Consider the following:
#generating some sample data that's fairly large
sample.data <- data.frame('target' = sample(c(1:10), size = 5000000, replace = T),
'regressor1' = rnorm(5000000),
'regressor2' = rnorm(5000000),
'regressor3' = rnorm(5000000),
'regressor4' = rnorm(5000000),
'regressor5' = rnorm(5000000),
'regressor6' = rnorm(5000000),
'regressor7' = rnorm(5000000),
'regressor8' = rnorm(5000000),
'regressor9' = rnorm(5000000),
'regressor10' = rnorm(5000000))
#building a toy glm - this one is about 3.3 GB
lm.mod <- glm(sample.data, formula = target ~ ., family = gaussian)
#baseline predictions
lm.default.preds <- predict(lm.mod, sample.data)
#extracting coefficients
lm.co <- coefficients(lm.mod)
#applying coefficients to original data set by row and adding intercept
lightweight.preds <- lm.co[1] +
apply(sample.data[,2:ncol(sample.data)],
1,
FUN = function(x) sum(x * lm.co[2:length(lm.co)]))
#clearing names from vector for comparison
names(lm.default.preds) <- NULL
#taa daa
all.equal(lm.default.preds, lightweight.preds)
Then we can do the following:
#saving for our example and starting timing
saveRDS(lm.co, file = 'myfile.RDS')
start.time <- Sys.time()
#reading from file
coefs.from.file <- readRDS('myfile.RDS')
#scoring function
light.scoring <- function(coeff, new.data) {
prediction <- coeff[1] + sum(coeff[2:length(coeff)] * new.data)
names(prediction) <- NULL
return(prediction)
}
#same as before
light.scoring(coefs.from.file, sample.data[1, 2:11])
#~.03 seconds on my machine
Sys.time() - start.time