Commit 407773a

| `spark.graphframes.useLocalCheckpoints` | Tells the connected components algorithm to use local checkpoints. If set to "true", the algorithm checkpoints to the local disks of the executors instead of to persistent storage. For "two_phase" this removes the need to configure a checkpoint directory. "randomized_contraction" always needs one, because it stores its intermediate tables there. Local checkpoints are faster but less reliable: they do not survive the loss of an executor. Used by the "two_phase" and "randomized_contraction" algorithms. | Optional (default: `false`) | 0.9.3 |

| `spark.graphframes.connectedComponents.checkpointinterval` | Sets checkpoint interval in terms of number of iterations. Checkpointing regularly helps recover from failures, clean shuffle files, shorten the lineage of the computation graph, and reduce the complexity of plan optimization. As of Spark 2.0, the complexity of plan optimization would grow exponentially without checkpointing. Hence, disabling or setting longer-than-default checkpoint intervals are not recommended. Checkpoint data is saved under the directory set by `org.apache.spark.SparkContext.setCheckpointDir`, or, when that is not set, under the `spark.checkpoint.dir` configuration; the algorithm writes into a sub-directory prefixed with its own name ("connected-components" or "randomized-contraction"). If neither is set, "randomized_contraction" throws a `java.io.IOException`, and so does "two_phase" unless local checkpoints are enabled. Set a nonpositive value to disable checkpointing. Used by the "two_phase" and "randomized_contraction" algorithms; note that "two_phase" in AQE mode (`broadcastthreshold` set to `-1`) relies on Spark's own DataFrame checkpointing instead of its own sub-directory. | Optional (default: `2`) | 0.9.0 |

- **Checkpoint Directory**: For configurations related to checkpointing, make sure to set a checkpoint directory using `spark.sparkContext.setCheckpointDir("path/to/checkpoint/dir")` before running algorithms that use checkpointing. Where `sparkContext` is not reachable, for example from a Spark Connect client, set the `spark.checkpoint.dir` configuration instead; GraphFrames falls back to it. For "two_phase" you can instead enable `spark.graphframes.useLocalCheckpoints`, which needs no checkpoint directory. "randomized_contraction" always needs one.

- **Local Checkpoints**: Local checkpoints are faster and can put strain on the local disk if insufficiently large. They are also less reliable than checkpoints in persistent storage, because they do not survive the loss of an executor. Because local checkpoints do not require setting `checkpointDir` for "two_phase", they are a convenient option for shorter runs.