The current trend in IT industry, especially in IoT world is to store all the data that is produced by all the devices or systems that are connected to a network. The storage is so cheap, that companies prefers to store and archive all this data, without caring if the data can be used now or in the future.
The main scope of this approach is to have all the information that might be needed in the future, if a system like Machine Learning or Hadoop is used. You never know what parameter or logs might become relevant in the future for an insight or a trend.
Things are becoming interesting in the moment when your devices contains more than100 or 200 sensors per device, with a sample rate of 1s or 0.1s.
Let’s take as example a bottle manufacture, that has plants all around the globe. Each plant has 800-1000 devices connected that runs 24/7.
The quantity of data that is produced every day can reach easily 0.2-0.5TB. Each day new data will arrive to the platform in the warehouse. Collected data might be processed and stored for later use. But, not all the time all the data is processed. You might want to keep data for later use for systems like Machine Learning.
The thing that should pop up in our mind is:
Do we need to store all the data that is produced by a plant?
At machine and gateway level, it might be very important to collect metrics at 0.01s or 0.05s time interval. But at plant level, this can be irrelevant. At plant level 0.5s or 1s time interval can be more than we need. This means that we already reduce the size of the data that is produced by a device from a plant with a factor of 5-10x.
When we look at global level, the information that are produced by devices are not relevant at second level. Not only this, but not all metrics are useful outside the plant. This means that from 100 or 200 counters we can end up with only half of them that are relevant, at global level.
This means that the data that we store at global level decrease drastically, but in the same time, we have all the relevant information that we might use in the future.
This approach doesn’t mean that we need at plant level complex processing system. On-premises or cloud, the system will look the same. It is our decision when and where we want to store data - at gateway, plant or global level. We could even have the gateway in cloud.
As an example we could have all the information produced by devices at gateway level stored for 7 days at device level. All the information that is older than 7 days is deleted automatically. The gateway can have a virtual plant in the cloud, where all data are stored for 1 year. At this two level the time interval of collected metrics is very low (0.1s and 1s). All this information is moved in the global repository that stores only a part of the counters at 1s time interval. This data can be later used for analytics and prediction.
As we can see we could have different approaches. Not all the time there is a need to store all the data that are produced by the devices forever. We can filter or decide what is the sample rate for each counter.
The main scope of this approach is to have all the information that might be needed in the future, if a system like Machine Learning or Hadoop is used. You never know what parameter or logs might become relevant in the future for an insight or a trend.
Things are becoming interesting in the moment when your devices contains more than100 or 200 sensors per device, with a sample rate of 1s or 0.1s.
Let’s take as example a bottle manufacture, that has plants all around the globe. Each plant has 800-1000 devices connected that runs 24/7.
The quantity of data that is produced every day can reach easily 0.2-0.5TB. Each day new data will arrive to the platform in the warehouse. Collected data might be processed and stored for later use. But, not all the time all the data is processed. You might want to keep data for later use for systems like Machine Learning.
The thing that should pop up in our mind is:
Do we need to store all the data that is produced by a plant?
At machine and gateway level, it might be very important to collect metrics at 0.01s or 0.05s time interval. But at plant level, this can be irrelevant. At plant level 0.5s or 1s time interval can be more than we need. This means that we already reduce the size of the data that is produced by a device from a plant with a factor of 5-10x.
When we look at global level, the information that are produced by devices are not relevant at second level. Not only this, but not all metrics are useful outside the plant. This means that from 100 or 200 counters we can end up with only half of them that are relevant, at global level.
This means that the data that we store at global level decrease drastically, but in the same time, we have all the relevant information that we might use in the future.
This approach doesn’t mean that we need at plant level complex processing system. On-premises or cloud, the system will look the same. It is our decision when and where we want to store data - at gateway, plant or global level. We could even have the gateway in cloud.
As an example we could have all the information produced by devices at gateway level stored for 7 days at device level. All the information that is older than 7 days is deleted automatically. The gateway can have a virtual plant in the cloud, where all data are stored for 1 year. At this two level the time interval of collected metrics is very low (0.1s and 1s). All this information is moved in the global repository that stores only a part of the counters at 1s time interval. This data can be later used for analytics and prediction.
As we can see we could have different approaches. Not all the time there is a need to store all the data that are produced by the devices forever. We can filter or decide what is the sample rate for each counter.
Wow - first time when cloud is optimized. Keep going!
ReplyDelete