Actual question
How can I serialize objects to ASCII and unserialize them again from ASCII without having to write to and read from a file connection (i.e. from ASCII that is in-memory)?
Background
In a state-less client-server framework, I would like to make certain information persistent accross calls (serialize >> send to client >> get serialized info back from client >> unserialize) without caching it on the server side.
Note that my JSON object/strong also contains other unserialized information and is thus mixed with the serialized information which is why the approach explained in this post doesn't completely do the trick.
Now, the thing is that I would like to unserialize the object solely based on the already-read JSON string. So to speak: from "in-memory ASCII" instead of from a file connection. How would I do that?
Here's what I tried:
require(forecast)
Approach 1
## SERVER: estimates initial model and writes JSON to socket
model <- auto.arima(AirPassengers, trace = TRUE)
## Model trace:
# ARIMA(2,1,2)(1,1,1)[12] : Inf
# ARIMA(0,1,0)(0,1,0)[12] : 967.6773
# ARIMA(1,1,0)(1,1,0)[12] : 965.4487
# ARIMA(0,1,1)(0,1,1)[12] : 957.1797
# ARIMA(0,1,1)(1,1,1)[12] : 963.5291
# ARIMA(0,1,1)(0,1,0)[12] : 956.7848
# ARIMA(1,1,1)(0,1,0)[12] : 959.4575
# ARIMA(0,1,2)(0,1,0)[12] : 958.8701
# ARIMA(1,1,2)(0,1,0)[12] : 961.3943
# ARIMA(0,1,1)(0,1,0)[12] : 956.7848
# ARIMA(0,1,1)(1,1,0)[12] : 964.7139
#
# Best model: ARIMA(0,1,1)(0,1,0)[12]
fc <- as.data.frame(forecast(model))
deparsed <- deparse(model)
json_out <- list(data = AirPassengers, model = deparsed, fc = fc)
json_out <- jsonlite::toJSON(json_out)
## CLIENT: keeps estimated model, updates data, writes to socket
json_in <- jsonlite::fromJSON(json_out)
json_in$data <- window(AirPassengers, end = 1949 + (1/12 * 14))
## SERVER: reads new JSON and applies model to new data
data <- json_in$data
model_0 <- json_in$model
model_1 <- eval(parse(text = model_0))
## Model trace:
# ARIMA(2,1,2)(1,1,1)[12] : Inf
# ARIMA(0,1,0)(0,1,0)[12] : 967.6773
# ARIMA(1,1,0)(1,1,0)[12] : 965.4487
# ARIMA(0,1,1)(0,1,1)[12] : 957.1797
# ARIMA(0,1,1)(1,1,1)[12] : 963.5291
# ARIMA(0,1,1)(0,1,0)[12] : 956.7848
# ARIMA(1,1,1)(0,1,0)[12] : 959.4575
# ARIMA(0,1,2)(0,1,0)[12] : 958.8701
# ARIMA(1,1,2)(0,1,0)[12] : 961.3943
# ARIMA(0,1,1)(0,1,0)[12] : 956.7848
# ARIMA(0,1,1)(1,1,0)[12] : 964.7139
#
# Best model: ARIMA(0,1,1)(0,1,0)[12]
# Warning message:
# In auto.arima(x = structure(list(x = structure(c(112, 118, 132, :
# Unable to fit final model using maximum likelihood. AIC value approximated
fc <- as.data.frame(forecast(Arima(data, model = model_1)))
## And so on ...
That works, but note that eval(parse(text = json_in$model)) actually re-runs the call to auto.arima() instead of just re-establishing/unserializing the object (note the trace information printed to the console that I included as comments).
That's not completely what I want as simply want to re-establish the final model object in the fastest possible way.
That's why I turned toserialize() next.
Approach 2
## SERVER: estimates initial model and writes JSON to socket
model <- auto.arima(AirPassengers, trace = TRUE)
fc <- as.data.frame(forecast(model))
serialized <- serialize(model, NULL)
class(serialized)
json_out <- list(data = AirPassengers, model = serialized, fc = fc)
json_out <- jsonlite::toJSON(json_out)
## CLIENT: keeps estimated model, updates data, writes to socket
json_in <- jsonlite::fromJSON(json_out)
json_in$data <- window(AirPassengers, end = 1949 + (1/12 * 14))
## SERVER: reads new JSON and applies model to new data
data <- json_in$data
model_0 <- json_in$model
try(model_1 <- unserialize(model_0))
## --> error:
# Error in unserialize(model_0) :
# character vectors are no longer accepted by unserialize()
Unfortunately, function unserialize() expects a file connection instead of "plain ASCII".
So that's why I need to do the following workaround.
Approach 3
## SERVER: estimates initial model and writes JSON to socket
model <- auto.arima(AirPassengers, trace = TRUE)
fc <- as.data.frame(forecast(model))
con <- file("serialized", "w+")
serialize(model, con)
close(con)
json_out <- list(data = AirPassengers, model = "serialized", fc = fc)
json_out <- jsonlite::toJSON(json_out)
## CLIENT: keeps estimated model, updates data, writes to socket
json_in <- jsonlite::fromJSON(json_out)
json_in$data <- window(AirPassengers, end = 1949 + (1/12 * 14))
## SERVER: reads new JSON and applies model to new data
data <- json_in$data
model_0 <- json_in$model
con <- file(model_0, "r+")
model_1 <- unserialize(con)
close(con)
fc <- as.data.frame(forecast(Arima(data, model = model_1)))
## And so on ...
Unserializing works now without the actual auto.arima() call being re-evaluated. But it's against my state-less paradigm as now the actual information is cached on the server side instead of actually being sent via the JSON object/string.
