PVV: We have to extract the requirements for the parameter manager. See the presentation of Raffaele in 2014.
PH: The current OCDB latency easily goes to minutes. We use snapshots to speed-up the access.
Slide 14: TPC drift object is ~100Kb, it was used to measure the performance.
GE: Do you present the time for round trip? Yes.
VB: Was there any consistency check? Yes, it was on the client side.
The access to the DAQ cluster is needed for full scale tests. We have 45 PCs in the DAQ reference cluster. Uli should grant access to the cluster. The network is 1Gb, no high throughput. The access is via gateway. We have to schedule a meeting with Pierre and Uli.
BvH: We nave to define the benchmarks and the number of objects, number of consumers in each scenario.
BvH: Jacek is working on a manager to merge the QA information. We can reuse it to collect the CDB objects.
VB: We have to measure separately "Get" and "Put" with a set of realistic tests taking into account the update frequency.
BvH: The merging takes time. For the TPC objects see the presentation of Jens.
MK: HLT test shows that the latency is negligible. The TPC drift velocity calibration takes 5 min. There are some trivial issues, the performance can be improved to ~1 min. when we fill the TPC information from the ITS/TPC matching and from the geom. propagation into the histograms.
Serialisation: there was no particular reason to use protobufs. We will keep the flexibility and select the most suitable serialisation package. Still the serialisation item should be kept in the agenda.
BvH: How do we provide access to the conditions DB fro T0, T1 and T2? One possibility is to populate again OCDB. We also can use "big DB" like CMS and ATLAS. What will be the access interface? ALFA. See the load balancer slide 7.
BvH: We should get rid of the "historical naming".
MF: What are the plans for engine implementation? There will be no OCDB in the future.
VB: We have a lot of similarities with the configuration system: parameters are size of time frames, scalers, numbers, strings. We should reuse the parameter manager in the configuration system. The access patterns however are different: lower frequency, low latency, smaller objects.
BvH: Why did we select RIAC? This was the work of a summer student last year, but the change of the backend is very easy. When we proved a set of automatic benchmarks, we will compare different DBs. In most of the cases we will use router - dealer pattern for load balancing (ZeroMQ).
Platforms: Linux, Mac
Feedback is needed.
MF: Would you align this tutorial with the official documentation? Yes, a link from README.md will be added after some testing.
AR: It is good to expose the tutorial to the collaboration after the implementation of the new API in FairRoot devices. The suggestions so far are very easy to implement.
