For the people interested in this issue, please feel free to submit individual pull request per estimator to make the review process faster by focusing review discussions on specific estimators.

I'll work on this ! May take some time to understand memorymaps & transformers better

As there's quite a few estimators/transformers to work on, I'll let myself to join. I'll start with FactorAnalysis, for instance ;)

Is PLS under a refactor right now? @arthurmensch

Sorry. Misclick.

Incremental PCA could fix this at the same time as #5173 . I see the issue but it probably means copying data per minibatch which may slow things down. And I am pretty sure @lesteve was using read only memmaps for his processing

I'm new to scikit and would wish to contribute to this issue. I'll have to go through memory maps, estimators and transformers. I'll stick to NuSVC and will get to the end of the problem as soon as possible

FYI the list of estimators that are expected to do something inplace should be found via:

>>> import inspect
>>> from sklearn.utils.testing import all_estimators
>>> pprint([e for e in all_estimators()
...         if 'copy' in inspect.signature(e[1].__init__).parameters])
[('AffinityPropagation',
  <class 'sklearn.cluster.affinity_propagation_.AffinityPropagation'>),
 ('Binarizer', <class 'sklearn.preprocessing.data.Binarizer'>),
 ('Birch', <class 'sklearn.cluster.birch.Birch'>),
 ('CCA', <class 'sklearn.cross_decomposition.cca_.CCA'>),
 ('FactorAnalysis',
  <class 'sklearn.decomposition.factor_analysis.FactorAnalysis'>),
 ('Imputer', <class 'sklearn.preprocessing.imputation.Imputer'>),
 ('IncrementalPCA',
  <class 'sklearn.decomposition.incremental_pca.IncrementalPCA'>),
 ('MaxAbsScaler', <class 'sklearn.preprocessing.data.MaxAbsScaler'>),
 ('MinMaxScaler', <class 'sklearn.preprocessing.data.MinMaxScaler'>),
 ('Normalizer', <class 'sklearn.preprocessing.data.Normalizer'>),
 ('OrthogonalMatchingPursuitCV',
  <class 'sklearn.linear_model.omp.OrthogonalMatchingPursuitCV'>),
 ('PCA', <class 'sklearn.decomposition.pca.PCA'>),
 ('PLSCanonical', <class 'sklearn.cross_decomposition.pls_.PLSCanonical'>),
 ('PLSRegression', <class 'sklearn.cross_decomposition.pls_.PLSRegression'>),
 ('PLSSVD', <class 'sklearn.cross_decomposition.pls_.PLSSVD'>),
 ('RandomizedPCA', <class 'sklearn.decomposition.pca.RandomizedPCA'>),
 ('RobustScaler', <class 'sklearn.preprocessing.data.RobustScaler'>),
 ('StandardScaler', <class 'sklearn.preprocessing.data.StandardScaler'>)]

TODOs are:

• check that for all Estimators given a readonly array, fit doesn't raise (even if copy=False in the init)
  -> if a copy is eventually needed and the user specified copy=False, do we raise a warning ?
• remove explicit ValueError raise
• check that there is never a check_array(copy=self.copy) if copy is not useful (which means if no inplace transformation is executed)
• check (optionnally) that transform with readonly data and copy=False raise for Estimators that are doing inplace modifications -> we may not be able to check this ?

I understood what needs to be done and was working on robust scaler, and when I tried to import it, it gave me a import error but the class robust scaler does exist in data.py in sklearn.preprocessing. I tried to even run the plot_robust_scaling.py given in the examples folder. Could you let me know what to do?
I'm a new comer to scikit!

Shouldn't SparseCoder also be in the list to be considered ? See #5956

I am working on this again

Is anyone working on this? This seems pretty bad.

MiniBatchDictionaryLearning has the same issue.

We also need to test on sparse arrays, probably of multiple types (CSR, CSC should be enough?)

I confirm that MiniBatchDictionaryLearning has the same issue.

Is the resolution of this bug can change the behavior/results of the algorithm?

Is the resolution of this bug can change the behavior/results of the algorithm?

Sorry I can't parse this sentence.

Sorry if it was not clear.

What I meant is: Does a fix of this bug will change the results obtained from the algorithm?

@rousseau no.

I just encountered this bug in sklearn.metrics.pairwise, so it's not restricted to estimators.

In my case I was trying to compute pairwise manhattan distances on a big sparse matrix. If I give it n_jobs>1 it crashes with ValueError: buffer source array is read-only message.

I did the hacky thing of changing the lines in manhattan_distance to say copy=True) and now it's chugging along, but that's not a general solution.

joblib is behind all this of course but AFAIR, the idea was to modify scikit-learn code to avoid doing inplace modification by default (copy=True by default and expose a copy parameter if not present already), see #6614 (comment).

@jamestwebber I created a separate issue #7981.

I am interested to work on it. However I need some clarifications:
The goal is to:

1. Add a parameter copy if it wasn't already
2. Put True by default to the copy parameter if False before

A 3rd point was done in PR #5507:
3. Force to true (even if set to False) the copy parameter if the array is read-only
It was done by adding before each check_array call:

copy = not X.flags.writable or self.copy 

Is it necessary ?

Finally, is this PR #4807 important before doing modifications ?

I removed the easy tag. @Morikko my personal advice: maybe try to tackle simpler PRs first before coming back to this one.

ValueError: output array is read-only when using n_jobs > 1 with RandomizedLasso.

Hi.. the same problem with TrucatedSVD

I think we need to call this one a blocker and make a point of fixing it... I'm not sure I'm as optimistic as @arthurmensch about how easy they are to fix... I think it would be good to see an example of one fixed, so that contributors can follow suit on others.

I also think we should try to merge #4807 and make solving this a matter of removing estimators from a list of those failing...