What cross validation use with time series database? I'm using a regression tree to predict/forecast a daily bases data.
I'm wondering to use a cross validation to train and predict all the data.
What cross validation procedure I may use?
 A: You'll likely want some sort of Roll Forward Cross-Validation. Roll-forward works like regular cross validation, only it stipulates that the testing data will always be immediately following the training data. Think of testing known observations by applying the model to a subset of your data that ends at a certain point, and then comparing the model results to a subset of your data that begins immediately after the point you ended your training on. 
Robert Hyndman has an excellent blog post on implementation of this is R here. 
A: There are several good questions embedded in your post. Let me try to tackle them in turn:  


*

*Should I use cross-validation, and if so which procedure?
Using cross validation to tune your model and hyper-parameters is a good idea. If you are considering using Python for your project then Scikit-Learn is a popular machine-learning package that may be useful. You will find more specifics on how to use the cross-validation function in Scikit-Learn here:  Cross-Validation on SKLearn
Importantly, for time-series, you should use the correct roll-forward split of data. The SKLearn team kindly included a time-series splitter in the recent v0.18 of Scikit-Learn:  SKLearn Time Series cross-validator

*Should I use the model to train and predict all the data? 
Even if you use cross-validation, it is still advised that you split a portion of test data off before you start training and validation. You can then keep this data for final testing once you have created your model. More discussion here:  Should I split my data to train/test split or train/validation/test subset?
