分布式训练为什么出现 Variables not initialized？

hello，咨询一下，我们进行分布式训练时（1 PS + 2 Worker），跑的是官网ppi数据模型，启动worker-0启动后会立刻才是计算，如下：
==> /tmp/log.worker.0 <==
INFO:tensorflow:f1 = 0.40184546, loss = 0.69358194, step = 0
INFO:tensorflow:f1 = 0.4026365, loss = 0.5608437, step = 20 (1.201 sec)
INFO:tensorflow:f1 = 0.4041432, loss = 0.53496814, step = 40 (1.069 sec)
INFO:tensorflow:f1 = 0.4073638, loss = 0.5300987, step = 60 (1.115 sec)

但是，worker-1启动后，会有一段等待时间，然后才开始进行计算，具体日志如下：
INFO:tensorflow:Graph was finalized.
2019-04-23 03:45:56.103402: I tensorflow/core/distributed_runtime/master_session.cc:1192] Start master session 27c7c5a9b0b92ad1 with config: gpu_options { allow_growth: true }
INFO:tensorflow:Waiting for model to be ready. Ready_for_local_init_op: Variables not initialized: supervisedgraphsage_1/sageencoder_1/meanaggregator_1/dense_2/kernel, supervisedgraphsage_1/sageencoder_1/meanaggregator_1/dense_3/kernel, supervisedgraphsage_1/sageencoder_1/meanaggregator_2/dense_4/kernel, supervisedgraphsage_1/sageencoder_1/meanaggregator_2/dense_5/kernel, supervisedgraphsage_1/dense_1/kernel, supervisedgraphsage_1/dense_1/bias, global_step, beta1_power, beta2_power, supervisedgraphsage_1/sageencoder_1/meanaggregator_1/dense_2/kernel/Adam, supervisedgraphsage_1/sageencoder_1/meanaggregator_1/dense_2/kernel/Adam_1, supervisedgraphsage_1/sageencoder_1/meanaggregator_1/dense_3/kernel/Adam, supervisedgraphsage_1/sageencoder_1/meanaggregator_1/dense_3/kernel/Adam_1, supervisedgraphsage_1/sageencoder_1/meanaggregator_2/dense_4/kernel/Adam, supervisedgraphsage_1/sageencoder_1/meanaggregator_2/dense_4/kernel/Adam_1, supervisedgraphsage_1/sageencoder_1/meanaggregator_2/dense_5/kernel/Adam, supervisedgraphsage_1/sageencoder_1/meanaggregator_2/dense_5/kernel/Adam_1, supervisedgraphsage_1/dense_1/kernel/Adam, supervisedgraphsage_1/dense_1/kernel/Adam_1, supervisedgraphsage_1/dense_1/bias/Adam, supervisedgraphsage_1/dense_1/bias/Adam_1, num_finished_workers, ready: None

从日志看到worker-1貌似在等待模型变量初始化：
INFO:tensorflow:Waiting for model to be ready. Ready_for_local_init_op: Variables not initialized

请教一下，这个问题怎么解决呢？

Start by reproducing the official PPI model with one parameter server and two workers, then compare the worker-0 and worker-1 startup logs. Investigate the distributed initialization path responsible for “Waiting for model to be ready” and the listed uninitialized variables; done means the cause of the delay is identified and a verified resolution or documented limitation is provided.

Written by the indexing model from the issue text.