Parsing huge data coming from a URL

I need to parse the data coming from the URL which looks like this:

hasProcess=true
version=1
DATACENTER=abc
TotalNumberOfServers:4
primary:{0=1, 1=2, 2=1, 3=2, 4=1, 5=2, 6=1, 7=2, 8=1, 9=2, 10=1, 11=2, 12=1, 13=2}
secondary:{0=0, 1=0, 2=0, 3=1, 4=0, 5=0, 6=0, 7=1, 8=0, 9=0, 10=0, 11=1, 12=0, 13=0}
hosttomachine:{3=machineA, 2=machineB, 1=machineC, 4=machineD}
DATACENTER=pqr
TotalNumberOfServers:2
primary:{0=1, 1=2, 2=1, 3=2, 4=1, 5=2, 6=1, 7=2, 8=1, 9=2, 10=1, 11=2, 12=1, 13=2, 14=1}
secondary:{0=0, 1=0, 2=0, 3=1, 4=0, 5=0, 6=0, 7=1, 8=0, 9=0, 10=0, 11=1, 12=0, 13=0, 14=0}
hosttomachine:{1=machineP, 4=machineQ}
DATACENTER=tuv
TotalNumberOfServers:0
primary:{}
secondary:{}
hosttomachine:{}


After parsing the data I need to store each datacenter data in a Map like this:

HashMap<String, Map<Integer, String>> primaryData


For example, the Key of primaryData is abc and value is:

{0=1, 1=2, 2=1, 3=2, 4=1, 5=2, 6=1, 7=2, 8=1, 9=2, 10=1, 11=2, 12=1, 13=2}


which is for primary.

Similarly another Map for secondary for each datacenter:

HashMap<String, Map<Integer, String>> secondaryData


For example, the Key of secondaryData is abc and value is:

{0=0, 1=0, 2=0, 3=1, 4=0, 5=0, 6=0, 7=1, 8=0, 9=0, 10=0, 11=1, 12=0, 13=0}


which is for secondary.

And lastly, one more map for hosttomachine mapping for each datacenter:

HashMap<String, Map<Integer, String>> hostMachineMapping -


For example, the Key of hostMachineMapping is abc and value is:

{3=machineA, 2=machineB, 1=machineC, 4=machineD}


which is for hosttomachine.

And all the above map will have data for its datacenter as I have three datacenter in the above example. So each each map will have three data. And also I will parse the above response only when hasProcess is equal to true. If it is not true, then I won't parse anything.

This code takes more than 200 ms to parse the data and store it in its corresponding HashMap. Is there any way to parse the above data efficiently and store it in particular HashMap?

private void parseResponse(String response) throws Exception {
if (response != null) {
Map<String, Map<Integer, String>> primaryData = null;
Map<String, Map<Integer, String>> secondaryData = null;
Map<String, Map<Integer, String>> hostMachineMapping = null;

long version = 0L;
boolean changed = false;
String splitResponse[] = response.split("DATACENTER=");

boolean flag = false;
for (String sr : splitResponse) {
if (!flag) {
flag = true;
if (!changed) {
return;
} else {
primaryData = new HashMap<String, Map<Integer, String>>();
secondaryData = new HashMap<String, Map<Integer, String>>();
hostMachineMapping = new HashMap<String, Map<Integer, String>>();
}
} else {
generateDataCenterMapping(sr, primaryData, secondaryData, hostMachineMapping);
}
}

if (changed) {
Mapping.setPrimaryData(primaryData);
Mapping.setSecondaryData(secondaryData);
Mapping.setHostMachineMapping(hostMachineMapping);
Mapping.setVersion(version);
}
}
}

private void generateDataCenterMapping(String sr, Map<String, Map<Integer, String>> primaryData,
Map<String, Map<Integer, String>> secondaryData,
Map<String, Map<Integer, String>> hostMachineMapping) throws Exception {

String[] data = sr.split("\n\t");
String dcName = data[0];
int numOfServers = Integer.parseInt(data[1].split(":")[1]);
if (numOfServers > 0) {
primaryData.put(dcName, generateMap(data[2]));
secondaryData.put(dcName, generateMap(data[3]));
hostMachineMapping.put(dcName, generateMap(data[4]));
}
}

private Map<Integer, String> generateMap(String map) throws Exception {
String tableString = map.split(":")[1];
Map<Integer, String> table = new HashMap<Integer, String>();
tableString = tableString.substring(1, tableString.length() - 1);
String[] entries = tableString.split(", ");

for (String e : entries) {
String[] entryVal = e.split("=");
table.put(Integer.parseInt(entryVal[0]), entryVal[1]);
}

return table;
}

• What's the name of the class with this parseResponse method? Does it not have any public members? – Mathieu Guindon Jun 23 '15 at 0:59
• Since your code is not about the getting of the data, then it actually does not matter where it comes from, so that detail should not be included prominently in the question title. – JK01 Jun 23 '15 at 2:45
• This data format looks like yaml, but it isn't. Do you also have control over the server, and could you make it emit YAML instead? – 200_success Jun 23 '15 at 10:03

I started to write a comment, but it became too long.

Questions

• 200 ms is pretty close to eternity, how long is your input?
• Can you measure how long parseResponse alone takes?
• Could you post your data? Someone may try harder to optimize.

Potential slowness causes

I guess your String.split could be the culprit (with more than one character a regex gets created; you should use

 private static final Pattern COMMA_BLANK_PATTERN = Pattern.compile(", ");


for splitting and use split limit of 2. Probably faster could be using Guava's

private static final Splitter COMMA_BLANK_SPLITTER = Splitter.on(", ");


Even better would be piece-wise processing instead of splitting.

Note also that Integer.parseInt is internationalized and therefore slightly slower than necessary. Probably unimportant.

Review

if (response != null) {


As Mat's Mug said, this is wrong. I'm using Guava's Preconditions with a static import writing simply

checkNotNull(response);


This makes it fail-fast rather than hiding the problem to bite you later.

String tableString = map.split(":")[1];
Map<Integer, String> table = new HashMap<Integer, String>();
tableString = tableString.substring(1, tableString.length() - 1);


This is pretty confusing by squeezing something else between the two tableString defining lines.

Summary

Overall, it's not bad. When speed is important, I'd go for something like

MyParser parser = new MyParser(response).skip("hasProcess=");
parser.skip("\nversion=");
while (parser.skipIfLookingAt("\nDATACENTER=")) {
parser.parseDatacenter(parser, whatever...);
}


It's just an idea (damn similar to java.util.Scanner, which isn't exactly known for speed), but I guess that string splitting is the problem (simply as I can't see anything else).

An example method could look like

MyParser skip(String prefix) {
for (int i=0; i<prefix.length(); ++i) {
if (content.charAt(index++) != prefix.charAt(i)) {
throw ...
}
}
}


where content and index are instance variables. This part doesn't need any substring creation at all (and gets a bit unreadable because of this). Warning: Doing something like content = content.substring(index) (so you could use startsWith) would make it clearer but terribly slow as you'd copy a big part of the string a lot of times.

Filling the maps

Now, I'd probably move all the functionality into MyParser, so that the parsing would be just

 private void parseResponse(String response) {
new MyParser(response).parse();
}


The parser would have fields like

 private final Map<String, Map<Integer, String>> primaryData;


so that I could write

void parseDatacenter() {
skip("\n\tTotalNumberOfServers:");
skip("\n\tprimary{");
skip("\n\t");
}

Map<String, String> result = new HashMap<>();
while (true) {
map.put(key, value);
if (lookingAt('}') {
break;
}
skip(", ");
}
return result;
}


As an example see this simple method

String readTillOneOf(String terminators) {
int start = index;
while (terminators.indexOf(content.charAt(index)) > -1) {
++index;
}
return content.substring(start, index);
}


I ignored quite some details like parsing numbers and I can't tell how many parsing methods you'll need, but basically it's all pretty simple. Just skip what's fixed and process what you want while looking for a terminator.

It could get a bit more complicated if you needed to be error-tolerant, like allowing multiple spaces, but then you could let some of your methods skip over them. Or, if it gets difficult, use regexes. But I don't think it gets complicated.

Note that I'm unsure why your parser is slow. There's a lot of splitting and this is the probable cause, but it's just a guess.

• Thanks for the help. Your approach is looking good to me. But I am confuse how do I adapt this to get the actual working solution. If possible and if you have some time, can you provide an example how will this work with your overall approach? – david Jun 23 '15 at 3:16
• @david I don't know good enough what you're doing. I'd just make the three "major" maps to instance variables and let parseDatacenter create and fill the per-datacenter maps and put them into the major ones. It should eat the string (i.e. increment index) so that the next iteration looks at "\nDATACENTER" again. +++ I can write some more code, but chose what part would be most helpful. – maaartinus Jun 23 '15 at 3:32
• To me important part is, given a string response like as shown in the question, what is the best way to populate those three maps? Since this will set the foundation for me to move forward. If you were to solve this problem, then how would you do this, your approach sounds good to me but I was confuse to adapt it to populate three maps. – david Jun 23 '15 at 4:09
• @david I've updated the answer. Basically, you need a parser for the map data, which starts just after the opening brace and exits on the closing brace. Then you put the map where it belongs to, eat the brace and other fixed things, etc. – maaartinus Jun 23 '15 at 6:06

I find it awkward that you're wrapping the entire method's body in an if block:

if (response != null) {


In fact, I find it awkward that you're doing nothing with a null response, especially given an indicator such as throws Exception.

I would expect the method to throw some kind of ArgumentNullException (sorry if that's not in Java, I'm a C# guy - I'm sure Java has something similar though) given a null response string.

The typical way to do this, is to have a guard clause at the start of the method:

if (response == null) {
throw new ArgumentNullException("response");
}


And then, go on with the method body, knowing you have a valid response. Notice that just eliminated a whole nesting level across the entire method!

I find your logic rather convoluted, and I would expect parseResponse to return an object (or a collection of objects) holding the parsed data... not to have direct (or indirect) side-effects: I think your code is doing too many things.

The fact that parseResponse isn't a public member smells IMO, because parsing a response is a testable concern of its own, and there should be a class dedicated to making this work, where parseResponse would be public. Having it buried as an internal implementation detail of something bigger tells me that the only way you have of testing (and benchmarking) your parsing logic is by running the entire thing - presumably without any way of really knowing how much of the 200ms is actually spent sending the request and receiving the response: you want to be able to feed that method with a "fake" response string, and test-run all edge cases without having to send an actual request to some remote server that could be offline for maintenance while you're running your tests!

• @Mat Thanks for your suggestion. This class is my background thread which runs every 5 minute and it calls another service url from which it gets the above data. Now after parsing that string, I need to set those three maps in my Mapping class so that's why it was void. – david Jun 23 '15 at 1:32

Unfortunately, there is quite a bit of fuzziness with respect to the performance issues here. First, what is actually taking 200ms? Is it just this method, exclusively? Is it this method plus the external methods it calls (the Mapping.set* methods)? Is it the entire request and parsing? Second, has there been any profiling done to narrow down some candidates for what exactly might be the bottleneck? It could be the string manipulation, the HashMap, or something else completely. Third, what kind of input are we looking at? There is a sample, but are there hundreds of datacenters or hundreds of thousands? If the number is truly large, 200ms may actually be completely reasonable. There is also the possibility of being bound by some component of the hardware, but we can hand wave that.

Given the constraints listed, here are a few suggestions. They are not tested because of the potential variability of the above. The shape and size of your actual data could lead to significant differences between the suggestions.

Strings

There is obviously a lot of string manipulation going on, particularly String#split. maaartinus goes into some detail on why this can lead to performance troubles, especially if it is using regexes internally. In place of some of the splits, String#substring, String#startsWith, and String#endsWith might be more performant.

It might also be faster to handle lines individually using a combination of StringReader and BufferedReader, or Scanner. Each of these has their own set of performance properties and optimizations to consider. These could even make your code slower depending on how they handle things internally, but it might be worth trying them.

HashMaps

HashMap is generally a great default choice for a map in Java. It provides O(1) insertion and retrieval in most cases. But it does have some quirks. First, if you are putting a lot of data in sequentially, the HashMap internally needs to keep growing its internal structures to keep up. This could cause performance problems because each time it grows, it needs to allocate a new chunk of memory, then rehash all the keys to put them into the new chunk. We could instead just tell HashMap what we would like the initial capacity to be, accounting for the load factor, so that this is not an issue. I have seen data to suggest that this may not be as big of an issue as suggested in the documentation, though.

Collisions can also affect the performance of a HashMap, but you are using built-in Java objects for keys. Generally, the built-in hashing algorithms produce few collisions on the general set of inputs.

The hashing algorithms themselves, though, may be a bottleneck depending on the shape of your data (the length of the keys). The built-in hashing algorithms are designed to be cheap, but they are not free. There is a possibility that the actual hashing is what is taking so long. This is doubtful, as the algorithms are cheap and the results can be cached, but it is a possibility.

Since you did not post what Mapping.set* does, I will make a suggestion based on a potentiality. If you are iterating over the HashMaps in those methods, you could be experiencing your performance losses in there. Iterating over a HashMap is dependent on the capacity rather than the number of elements. A HashMap with a capacity much larger that the number of elements will slow down.

You can try to vary the type of Map you use to see if one of these is affecting the performance. For example, you could use a LinkedHashMap if iteration is the problem. You could test with a TreeMap to see if a completely different structure may actually be more suitable in practice. Because you have two levels of Map with different key types, they may each benefit from different optimizations.

Other things

I agree with @Mat's Mug about using guard clauses to reduce the indentation, and I see more places where similar strategies may be appropriate.

I see you have a variable name flag, and I don't like that. It's used awkwardly, and the fact that it has such a nondescript name tells me that it doesn't have a good purpose. Let's look at what it actually does:

   boolean flag = false;
for (String sr : splitResponse) {
if (!flag) {
flag = true;
if (!changed) {
return;
} else {
primaryData = new HashMap<String, Map<Integer, String>>();
secondaryData = new HashMap<String, Map<Integer, String>>();
hostMachineMapping = new HashMap<String, Map<Integer, String>>();
}
} else {
generateDataCenterMapping(sr, primaryData, secondaryData, hostMachineMapping);
}
}


It is used in the loop to switch behavior. The conditional is a negated conditional with an else clause, which is kind of a smell. Generally, I would start by reversing the conditional, but here there's more. Immediately after the conditional, flag is set to true, and then never set again. I see what this is. This is a "just run me the first time" chunk of code. We can take this out, and just adjust the loop to skip the first iteration.

if(splitResponse.length == 0) {
return;
}

if (!changed) {
return;
}

primaryData = new HashMap<>();
secondaryData = new HashMap<>();
hostMachineMapping = new HashMap<>();

for (int i = 1; i < splitResponse.length; i++) {
String sr = splitResponse[i];
generateDataCenterMapping(sr, primaryData, secondaryData, hostMachineMapping);
}


I pulled that code out, but I also did a few other things. Because the code is now outside the loop, I had to add a guard to make sure that there is at least one element in the splitResponse array. Since the conditional on !changed returned, the else was superfluous, so I just removed it and un-tabbed the code inside. I also abbreviated the HashMap instantiations with the diamond operator(Java 7+). The loop is much simpler and the boolean flag is completely gone.

The changed flag can also be eliminated through similar steps.

I might also suggest (though the changes would be more significant) that, instead of having three separate Maps, you could create an object to hold the three pieces of data for each key, then have a single map from the key to the object. This would make the association between those data more explicit.