I found the problem.
I mixed curator and ILM for index management.
An ILM policy has crashed because of a closed index and grown up until 673Gb.
It started to saturate memory JVM. I reopened closed index and retry ILM policy. This solved ILM problem and a new index was created.
But this was not enough. The cache had to be clean to allow reassigning. After a cleaning and a rerouting process, the cluster is now green.
Here are the command which help me to solve it 
:
Get allocation explanation
GET /index-000170/_ilm/explain
Open closed index
POST /index-000143/_open
Retry ilm policy :
POST /index-000170/_ilm/retry
Get stat of fieldata :
GET _stats/fielddata?fielddata_fields=*
Clear cache fielddata :
POST /index-00016*/_cache/clear
Restart failed allocation
POST _cluster/reroute?retry_failed=true
Sources :
I hope this can help someone a day.