When running on CircleCI, different tests are found to be flaky in different test environments. Here are the details:
Test environment
Host OS on CircleCI: Ubuntu 14.04 (kernel: 3.13.0-151-generic) → Tests are flaky
(As comparison, on Jenkins we use Ubuntu 16.04 (kernel: 4.4.0-1062-aws) → Tests are stable)
OS in Docker container: Ubuntu 14.04 (kernel: 4.4.0-1062-aws)
How are we sharing the build products from build stage to test stage?
Build products are all in the intermediate Docker image which is shared from build stage to test stage.
Flaky tests
- TestFcOperator.test_fc
- caffe2_py2_gcc4_8_ubuntu14_04_test, https://circleci.com/gh/pytorch/pytorch/40627
- TestCRFOp.test_crf_gradient
- caffe2_py2_gcc4_8_ubuntu14_04_test, https://circleci.com/gh/pytorch/pytorch/40508
- TestReduceOps.test_reduce_min
- caffe2_py2_gcc4_8_ubuntu14_04_test, https://circleci.com/gh/pytorch/pytorch/40506
- TestReduceFrontSum.test_col2im_gradients
- caffe2_py2_gcc4_8_ubuntu14_04_test, https://circleci.com/gh/pytorch/pytorch/40771
- TestLayerNormOp.test_layer_norm_op
- caffe2_py2_gcc4_8_ubuntu14_04_test, https://circleci.com/gh/pytorch/pytorch/40507
- caffe2_py2_gcc4_8_ubuntu14_04_test, https://circleci.com/gh/pytorch/pytorch/40972
- TestLayerNormOp.test_layer_norm_grad_op
- caffe2_py2_gcc4_8_ubuntu14_04_test, https://circleci.com/gh/pytorch/pytorch/41306
- RecurrentNetworkTest.test_sum_mul
- caffe2_py2_gcc4_8_ubuntu14_04_test, https://circleci.com/gh/pytorch/pytorch/41027
- TestAdagrad.test_row_wise_sparse_adagrad
- caffe2_py2_gcc4_8_ubuntu14_04_test, https://circleci.com/gh/pytorch/pytorch/41227
Test environment
Host OS on CircleCI: Ubuntu 14.04 (kernel: 3.13.0-151-generic) → Tests are flaky
(As comparison, on Jenkins we use Ubuntu 16.04 (kernel: 4.4.0-1062-aws) → Tests are stable)
OS in Docker container: Ubuntu 16.04 (kernel: 4.4.0-1062-aws)
How are we sharing the build products from build stage to test stage?
Build products are copied from build stage to test stage, without sharing the intermediate Docker image.
Flaky tests
- GRUCellTest.test_gru_unit_op
- caffe2_py2_cuda9_1_cudnn7_ubuntu16_04_test, https://circleci.com/gh/pytorch/pytorch/39963
- TestLayerNormOp.test_layer_norm_grad_op
- caffe2_py2_cuda8_0_cudnn6_ubuntu16_04_test, https://circleci.com/gh/pytorch/pytorch/39939
- TestGlu.test_glu_old
- caffe2_py2_cuda8_0_cudnn6_ubuntu16_04_test, https://circleci.com/gh/pytorch/pytorch/40017
- caffe2_py2_cuda9_1_cudnn7_ubuntu16_04_test, https://circleci.com/gh/pytorch/pytorch/39980
- TestConvolution.test_conv_separate_stride_pad_gradients
- caffe2_py2_cuda9_1_cudnn7_ubuntu16_04_test, https://circleci.com/gh/pytorch/pytorch/39989
- TestReduceFrontSum.test_col2im_gradients
- caffe2_py2_cuda8_0_cudnn6_ubuntu16_04_test, https://circleci.com/gh/pytorch/pytorch/39979
- TestGroupConvolution.test_group_convolution
- caffe2_py2_cuda9_0_cudnn7_ubuntu16_04_test, https://circleci.com/gh/pytorch/pytorch/39978
When running on CircleCI, different tests are found to be flaky in different test environments. Here are the details:
Test environment
Host OS on CircleCI: Ubuntu 14.04 (kernel: 3.13.0-151-generic) → Tests are flaky
(As comparison, on Jenkins we use Ubuntu 16.04 (kernel: 4.4.0-1062-aws) → Tests are stable)
OS in Docker container: Ubuntu 14.04 (kernel: 4.4.0-1062-aws)
How are we sharing the build products from build stage to test stage?
Build products are all in the intermediate Docker image which is shared from build stage to test stage.
Flaky tests
Test environment
Host OS on CircleCI: Ubuntu 14.04 (kernel: 3.13.0-151-generic) → Tests are flaky
(As comparison, on Jenkins we use Ubuntu 16.04 (kernel: 4.4.0-1062-aws) → Tests are stable)
OS in Docker container: Ubuntu 16.04 (kernel: 4.4.0-1062-aws)
How are we sharing the build products from build stage to test stage?
Build products are copied from build stage to test stage, without sharing the intermediate Docker image.
Flaky tests