Index Based Salaries

Prices change. Some change every day, others may be once a month. Salaries are fixed, unless company decides to change it.

When government decides to increase petrol prices, price of almost everything increases because almost every business depends on transportation. But salaries don't change. When SBI increases cash reserve ratio or base rate, EMI increases, production decreases but salaries don't change.

It is kind of unfair that everyone in business can increase/decrease prices either based on market conditions or increase/decrease in cost of doing business but salaried people don't have that choice. The only choice they have is - talk to your manager or find a new job.

What if salary was a function? Say Hari Sadu has a home loan, drives a car and has children studying in school.  If his salary was linked to his home loan rate, cost of petrol and cost of education, he would be a very happy employee. Just like businesses can adjust their prices, if salaries could be adjusted the same way, it would create a much faster feedback loop for the economy to adjust to the new conditions. If a business decides that it cannot support salary function of some employee and wants to decrease his salary, it will come out clean and open. The current system of things, simply says things are same (salary stays the same) but actually employee is taking a hit and it stays unsaid, unacknowledged.  A 10% salary raise at yearly performance evaluation might be less that inflation or much much less than "impact of inflation" for a given employee.  In this new system, raise is a raise, it gives employee more spending power.

This is obviously a bad move for the companies because it makes one of fixed costs of business, variable. But what it also does is that it brings transparency. Just as businesses understand and adjust to prices of other businesses, it makes sense to understand and adjust to impact of prices on your employees.  If home loan rate increase by HDFC is eating into margins of Wipro or Infosys, they are in a better position of negotiate with HDFC than each employee on his own and then may be HDFC will figure out a deal where Wipro or Infosys charges less for software.  I guess all I am talking about is businesses being more people aware than money aware, because money in itself can do nothing, only people can.

Finding Keys In Cacheismo

In the last post I talked about how to do map-reduce like stuff with cacheimso. It occurred to me that their is no method to get all the keys. The global hashmap object now supports new method getPrefixMatchingKeys which as the name implies returns keys which match a certain prefix. Typically objects are stored with following template objectType$objectKey. Thus an object created by file quota.lua with key myKey will be stored with key quota$myKey. To do something with all the quota objects:
   
local keys = getHashMap():getPrefixMatchingKeys("quota$")
for k,v in pairs(keys) do 
   print(v)
end

If you would like the quota object itself to support say getAllKeys method then ...add this to quota.lua in scripts directory.



function getAllKeys() 

    local keys = getHashMap():getPrefixMatchingKeys("quota$")
    local result = ""
    for k,v in pairs(keys) do 
       result..v.."\n"  
    end
end

Now a get request with key quota:getAllKeys would return a new line separated list of all active quota objects in the map.

This is good enough but probably not very interesting. I am planning to support indexes as first class objects to make it faster to reach interesting objects in quick time. These indexes will automatically remove objects which are deleted or forced to expire because of LRU caching. So if you need to find all the quota objects that are close to quota limit, create an index on quota used value.

Using Cacheismo Cluster for Real Time Analytics

Cacheismo now supports invoking requests on other server.
  - getFromServer
    This function takes the server name and a key as an argument and returns the result to the script.
  - getInParallel
    This function takes map of server names to list of keys and gets values for all the keys from all the servers in parallel. Once all the results are received, the script gets a map of server names to map of keys and received values.

Here is a simple example code which gets to top accessed keys from the cacheismo cluster using a simplified, framework-less map-reduce.  The contents below should belong to file names mapreduce.lua

-- function called on every cacheismo node to get 
-- the top keys by access count
-- virtual key : mapreduce:topKeys:


function topKeys(count) 
    local keys  = accessCounter:getKeyAccessCount()
    -- sort keys by access count
    table.sort (keys)
    local result = {}
    local i      = 1
    for (i in 1..count) do 
       result[i] = keys[i]
    end
    return result
end


-- function called on one of the cacheismo nodes 
-- which distributes work to all other nodes using getInParallel
-- virtual key : mapreduce:manageGetTop:10 


function manageGetTop(count) 
     local query = {}
     for k,v in pairs(servers) do 
        query[k] = "mapreduce:topKeys:"..count
     end
     local results = getInParallel(query)
     local total = {}
     for k,v in pairs(results) do 
           for k1,v1 in pairs(v) do 
              total[k1] = v1
           end
     end
     table.sort (total)
     local topkeys = {}
     for (i in 1..count) do 
       topkeys[i] = total[i]
     end
end


I have taken some liberties with the syntax and their is no accessCounter object by default in cacheismo, but is fairly easy to create in the script.  Note that the above implementation doesn't have ant intermediate nodes, but is trivial to create by calling get for mapreduce:manageGetTop instead of calling mapreduce:topKeys.

Single cacheimso server might be serving thousands of such queries because everything is happening in non-blocking manner but we give user the illusion of synchronous stack using lua coroutines.

Links:
Cacheismo Introduction
Cacheismo Lua API
Cacheimso Sliding Window Counter Implementation in Lua
Cacheismo Quota Implementation in Lua
Cacheismo Memory Allocation
Cacheismo Source Code
Cacheismo Support

Multi-threaded vs Multi-process

Cacheismo doesn't use threads. So does redis. Memcached is multi-threaded. How to choose?
Multi-threaded applications are hard to code and understand. Non-blocking multi-threaded applications are close to nightmare. I had the honor of writing multi-threaded non-blocking http proxy @apigee. It was written in c, had JNI bindings with java and could talk to our custom interface with kernel using shared memory (to avoid kernel space to user space data copy).  Having done that, I choose not to use threads in cacheismo.

Scalability of multi-threaded applications with multiple cores is simply a function of how much work can be done in parallel. The lesser the conflicts between the threads, better is the performance. The first problem is distribution of work among the requests. If multiple threads are working  concurrently on same piece of data, it is hard for them to make progress. Memcached uses a single selector thread to bind connections to different threads. It is not optimal. It will become bottleneck when you have thousands of clients trying to connect to the server at the same time. May be all connections of one thread are sleeping whereas all connections of some other thread are killing the cpu. So even though we have some work, few threads are not bothered because it doesn't belongs to them. This is kind of not possible with memcached (it hardly uses cpu), but very much possible with cacheismo running user scripts.  Now consider the hashmap itself. One hashmap, multiple threads, unavoidable lock. Possible solution here would be striping. What about slabs. One possibility is per thread slabs or some form of thread local caching or instead of using lock at slab level, use per slab locks.

In short, to make multi-threaded code work at optimal levels, we have to looking at our objects in finer granularity than what we originally planned. We have to make our design sub optimal to make sure threads make progress, even if at the cost of under utilization of resources.
Multi-process model works the opposite way. Consider your complete code/process as an object and create multiples of them. Have some way to distribute requests to these individual objects/processes. In many cases it is possible, specifically when either clients have knowledge of servers (consistent hashing) or if some intelligent proxy can do this for us (http load balancer). Even this could be non-optimal, but at least it is simple. What else? Memory leaks are easier to manage with multi-process model. Just restart the process. Since memory is managed at process level, it can be easily reclaimed by killing the process and if we have multiple of them, the service is not disrupted for every user. In a multi-threaded model, restart will impact all users. No locks are needed in multi process code, unless we use shared memory. The process boundary gives exclusive access to memory which doesn't need further management.

I don't see multi-process code in any ways inferior to multi-threaded code. It has some benefits like simple to code, simple to debug, simple to maintain. Other beauty is if instead of running 1 process with multiple threads you are running multiple processes with single thread, crash in the first case will cause full system failure, whereas in second case you still have (n-1)/n percent of the system up.  It is easy to upgrade things one at a time because we have more than 1 thing. With single server, no option but to shut it down, upgrade and start.

The only problem I see with multi-process code is duplication of basic data structure and configuration.  For example running each requests in a separate java process in tomcat or jboss will be a really bad idea. There is simply too much shared state - db connection pool, byte code, runtime native code, session information, etc.   The other problem is a bit superficial one; instead of managing one process, we have to manage multiple of them (use scripts?)
If your solution is either stateless or it can be partitioned such that there is no interdependence, multiprocess will make more sense. It is a difficult choice, but just don't use threads if you can survive without them, specially with non-blocking code. Try  cacheismo and see for  yourself.

Cacheismo Cluster Update 2

Today I could set the value on server 1 and get it back from server 2.  This script of server 2 was changed a bit.


function handleGET(command, key) 
    local cacheItem = getHashMap():get(key)
    if (cacheItem ~= nil) then
   command:writeCacheItem(cacheItem)
   cacheItem:delete()
else 
   local result = command:getFromExternalServer("127.0.0.1:11212", key)
   if (result ~= nil) then
       writeStringAsValue(command, key, result)
   end 
    end
end

This is where I feel lua beats javascript because of its excellent coroutine support.  I could easily change the above code to a while loop and go through all servers in the cluster looking for a value and it would not change the complexity of the code a bit, but try doing it with javascript callbacks and it would look like a mess. 

I am thinking now that the original idea of using dataKey and virtualKey is a bit lame. What I would do now is that I will make server an explicit argument, just like I used above. For consistent hashing we will have an object exposed in lua which can be used to find the correct server via consistent hashing, but it would be optional to use. This gives users the flexibility to use some cacheismo instances as pure proxies and others as caches. Some of the servers can be part of one consistent hashing group and others can be part of another. For some reason if users want to use one of the servers as a naming server to map keys to server, they can use it that way. Essentially users have full flexibility with respect to how to map keys to servers  - consistent hashing, multiple servers hidden behind a proxy(cacheismo), naming server to map keys to servers or any combination of these.  

Other possibilities include replication and quorum. These can be accomplished with the current API itself, but I will add parallel get support to make it as fast as possible.