I think a standard based approach is the real answer to your question. Dealing with IoT we usually do not to focus on data meaning. This, IMHO, is still due to a deep human centric view of the problem: there always is a human figuring out what a sensor is reading. Usually this is done inferring information basing on environment, installation and so on, or assuming some lacking information when needed. Humans also apply a likeliness filtering on sensed information, usually being over confident on their own personal experience. Despite this could lead to devastating results, it also allows to make a better use of real time data.

If we will fail in providing "smart things" a way to understand each other we will obtain nothing better than what we have done in the real World, in which language is one of the first barriers we need to overcome in order to increase our access to informaion.

Implementing a standard way to share information should represent the foundation of IoT.  There are groups and task forces working on those topics. If you are interested have a look at IEEE P2413, to cite one.

I have to ask why you would like to index the real-time data? Would you perhaps benefit from structuring and classifying data into some sort of data model first, then consider how this data will be accessed. Once you can establish how the data will be accessed, you can then determine what indexes may be required to speed up access to the data for specific use cases.

You may also benefit from intermediary processing of some sort, if information is coming in many different formats. In this model the intermediate processing could structure and classify the data, then push it into any data structure you like. Also while the data is being structured and classified, responses can be built to respond to the real-time data events.

I want to answer some question like this 'find temperature sensors which currently read 33' or 'find quite rooms in building x', where temperature sensor may changes its readings frequently and room state was controlled by more than one sensor

I recently discovered this paper that suggests a 5-level semantic data model for IoT:

Sheth, A., Henson, C., & Sahoo, S. (2008), "Semantic Sensor Web," IEEE Internet Computing, July/August, p. 78-83. 


Perhaps this would help you develop your data model and identify where in the IoT data chain could be best to introduce the level of processing you require, to meet your goals of deriving intelligent action from real-time data.

Based on your questions, you may also want to consider a probabilistic approach. For example analysing historical data to determine the most likely rooms at 33, or most likely quietest rooms, based on the time, day, week, month, weather, etc. This approach could help generate a very likely answer without needing to analyse any real-time data at all :)